US 11,645,479 B1Grant
Method for AI language self-improvement agent using language modeling and tree search techniques
Issue Date:2023-05-09
•13 Claims
•18 Drawing Sheets
Abstract
A novel method provides an AI language virtual agent having self-improvement features and which uses language modeling and tree search techniques. The AI language virtual agent exchanges textual discussion with users and other simulated agents. The method includes receiving a current situational description depicting natural language user input, temperament qualities and textual tendencies of the virtual agent, and indicia regarding subject matter context of a present conversation. The indicia regarding subject matter context include textual logs from recent conversational exchanges. The current situational description includes audio, visual, and tactile inputs collected proximate to the virtual agent. The method preferably utilizes an MCTS tree search in combination with self-moving modules, one or more language models, tree search techniques outputting textual responses to the current situation description, and the virtual agent responding with textual expression to verbal input in combination with the audio, visual, tactile, and other sensory inputs.
Metadata
Inventor
- Kino High Coursey
Application Information
Application Number:US 17/093,608
Filing Date:2020-11-09
Priority Date:2019-11-07
Art Unit:2652
Classifications
IPC:
G06F40/58
Patent Drawings (18 sheets)
Description
Reference to Related Application
[0001] The present application is related as a continuation-in-part to U.S. Provisional Patent No. 62/931,815, filed Nov. 7, 2019, entitled “A Practical Method for Creating a Self-Improving Agent Using Language Modeling and Tree Search,” and invented by Kino Coursey.
Technical Field of the Invention
[0002] The present invention relates to artificial intelligence in language applications and in particular to machines programed to use artificial intelligence to simulate agents for interacting and communicating with humans.
Background of the Invention
[0003] Prior art speech recognition and speech synthesis computing has been provided by recognizing and responding to a list of predetermined prompts. Speech generated in response to speech recognition under various situations is substantially more complex than speech generated in response to the predetermined prompts. In the field of language generation, language models trained on large databases, such as OpenA1 GPT-2 and GPT-3 have produced near human quality output. In other fields such as AI board games, tree search-based playing of board games has resulted in systems such as AlphaZero developing high performance in a mostly self-taught manner through repeated self-play and logging of self-play games. Logging of self-play games creates self-play logs which are later used for tree search queries and predicting expected outcomes of various game moves in selecting the moves having the highest probability of game success. The prior art has not provided a highly successful virtual agent for language interactions which utilizes self-play learning to generate conversation logs from tree search processes in determining language utterances.
Summary of the Invention
[0004] A novel method for AI language self-improvement agent using language modeling and tree search techniques is disclosed for a virtual agent exchanging textual discussion with users and other simulated virtual agents. The method includes the steps of receiving a current situational description, wherein the current situational description includes natural language user input, properties regarding the qualities of the virtual agent, and indicia regarding subject matter context of a present conversation. The qualities of the virtual agent include temperament and textual tendencies. The indicia regarding subject matter context include textual logs from recent conversational exchanges. The current situational description includes audio, visual, and tactile inputs collected proximate to the virtual agent. The method utilizes a database of one or more language models, conversation logs storing text from prior textual exchanges, and reference conversations utilized for training according to the one or more language models. The method is comprised of steps for executing instructions for a combination of self-play engines for training of the language model with self-play and external interaction engines for communicating with one or more external users or external virtual agents. The method further includes instructional sets for self-moving modules for advancing the method of external agents/users communicating with the virtual agent via a combination of textual exchanges and one or more audio, visual or tactile inputs into the virtual agent. The method preferably utilizes tree search processes, such as Monte Carlo Tree Search (“MCTS”) processes, in combination with the one or more language models to provide the tree search techniques outputting textual responses to said current situation description, and wherein said virtual agent responds with textual expression to verbal input in combination with the audio, visual, tactile, and other sensory inputs.
[0005] A method is disclosed for embedding language models such as GPT-2 or GPT-3 within a tree search framework, allowing the benefits of both methods to be realized, and to provide additional benefits from their mutual interaction. The agent can select its utterance or action based on the projected expected outcome. The language model provides both “move generation” and “evaluation” functions for the search process. The search process simulates multiple viewpoints conducting a virtual conversation based on the language models used. The interaction between viewpoints is a deliberation process and is inspectable, allowing explanation of outcome. Self-play/Self-talk allows the system to generate self-play logs which are used to train future versions of the language model. Self-play logs are used to train the language models and “compile” overall system performance into the language model, improving both the move generation and evaluation process. Language models can generate many plausible responses, but can lack a way of selecting the “best” response versus the “most probable” response. The tree search process selects the “best” response leading to the highest average expected outcome based on sample-based projections of the future. Goal directed self-filtering is also provided in selecting appropriate choices.
Description of the Drawings
[0006] For a more complete understanding of the present invention and the advantages thereof, reference is now made to the following description taken in conjunction with the accompanying Drawings in which FIGS. 1 through 17 show various aspects for a method for AI language self-improvement agent using language modeling and tree search techniques according to the present disclosure, as set forth below:
[0007] FIG. 1 is a flow chart depicting the method for AI language self-improvement agent using language modeling and tree search techniques according to the present disclosure;
[0008] FIG. 2 is a block diagram depicting the overall architecture of a system implementing the AI language methods according to the present disclosure;
[0009] FIG. 3 is a flow chart depicting a process for generating a contextual description;
[0010] FIG. 4 is a flow chart depicting an abstracted tree search process utilizing contextual descriptions;
[0011] FIG. 5 is a schematic diagram depicting different layers of a Monte Carlo Tree Search Process;
[0012] FIG. 6 is a flow chart depicting a move generation process using language models;
[0013] FIG. 7 is a flow chart depicting an evaluation process using language models;
[0014] FIG. 8 is a flow chart depicting steps for evaluation generation;
[0015] FIG. 9 is a flow chart depicting evaluation utilizing a basic sentiment analysis;
[0016] FIG. 10 is a flow chart depicting evaluation using goal oriented analysis;
[0017] FIG. 11 is a flow chart which depicts merging multiple analyses into a weighted evaluation;
[0018] FIG. 12 is a schematic diagram depicting an apparatus for self-learning through application of a self-play feedback loop;
[0019] FIG. 13 is a schematic diagram depicting an apparatus for interlocking training and self-play processes;
[0020] FIG. 14 is a schematic diagram depicting a design for an agent for generating linguistic responses to verbal and situational inputs;
[0021] FIG. 15 is a flow chart depicting a training process required for a broad coverage agent;
[0022] FIG. 16 is a schematic diagram for providing an informational funnel to construct Immediate Situational Descriptions; and
[0023] FIG. 17 is a block diagram for generating and selecting content using entropy density and a fractional knapsack model.
Description of the Invention
[0024] FIG. 1 is a flow chart depicting a method for an AI language self-improvement virtual agent using language modeling and tree search techniques. Two interlocking processes are shown that together implement a method for generating natural language output and initiating actions through a deliberative process and a method of improved operation through both supervised and unsupervised learning. The deliberative interaction process is started in Step 12 and receives situational input from either internal or external sources in Step 001. Such description may include user input expressed in natural language optionally combined with sensory or situational input expressed also in natural language. In Step 002, the situational input is merged with background information to create a contextual description. The goal of the contextual description is to provide a summary of all relevant information relevant to a deliberation process other than the immediate situational input. In Step 003, a deliberative search process uses the contextual description to produce a range of possible actions and to assign values to each possible action. This search process uses language models 008 to both produce plausible continuations and to provide evaluative information. The output of this process is a list of possible actions along with their score or value. The result of the search process 003 will produce a data structure that encodes the expected values of each possible action under consideration. In Step 004, the list of possible actions and their evaluations are extracted from the search data structure. In Step 005, the optimal action is extracted from the list of actions and their evaluations. Usually, this selects either the maximum or minimum action based on the evaluation used; however, other criteria may be used and the value of each action may be a multi-dimensional vector instead of a unitary value, in which case a weighted selection process may be employed. In some cases, the action selected may not be optimal from a numeric view but may be selected to meet some other criteria such as variety or behavioral loop prevention. In Step 007, the action selected in Step 005 is implemented through interpretation as either a user interaction, animation action or as a system internal action. The output of Step 007 becomes part of the history of an interactions log that is used to make up the immediate situational input of Step 001.
[0025] In Step 006, the situational input and optionally contextual description is recorded along with the selected output of Step 005 into an interaction log 009. The interaction log is used by the unsupervised learning process system in Step 011. Once initiated in Step 14, the language model training process 011 generates the language models 008 from reference material 010 used for supervised training and interaction logs 009 for unsupervised learning. Reference material 010 may be provided by manually selected material, by curated material, or by automated collection methods, such as topic focused collection, quality focused web collection, and text mining. During the search process in Step 003, numerous simulated interactions may be generated, and these simulated interactions may form additional training material for unsupervised learning of both generative and evaluative language models or classifiers used in the system.
[0026] FIG. 2 is a block diagram depicting the overall architecture of a system implementing the AI language methods according to the present disclosure, and represents additional details for the process of both deliberative search and learning represented in FIG. 1 . An object 100 represents the human or universe interacting through the user interface 101. The external user interface 101 is the primary demarcation of the system and the outside world. The external user interface 101 collects any relevant information such as the user's 100 utterances, sense impressions from sensors 103 or parameters useful to control the search process or to provide context. The combination of processes 100 and 101 correspond to Step 001 in FIG. 1 . The user interface 101 provides a packet of information to the system in the form of a contextual description packet 102. This packet 102 includes information for the search settings, background information from sensors 103 to provide a description of the “scene” being imagined, the conversational history of the dialog or actions between the system and the human user 100, and conversational prompts given to the language model. A portion of packet 102 provides the search setting used by the MCTS processor 106. The other portion of packet 102 provides the conversational context and scene description which are passed to the contractual prompt generator 104 which generates the initial contextual prompt given to the MCTS processor 106 and is applied at the initial node or root of the MCTS processor 106. This is required to provide a description that fits within the processing capacity of any language model being used. The combination of processes 102 and 104 correspond to Step 002 in FIG. 1 .
[0027] The search process 106 uses the information provided by the root node generation process 104 and search settings from packet 102 to conduct a search for an action with the highest expected future reward from the set of possible initial actions 107, using best response extraction process 120. The search process 106 in the initial embodiment uses Monte Carlo Tree Search or MCTS. MCTS requires that the state description encapsulated by each search node provide four functions: “getPossibleAction( )” returning a list of actions (conversational or otherwise) that can be taken from that state, “takeAction(action)” which returns the new state from taking that action; “isTerminal Q” which reports if a state is a terminal or leaf state; and “getRewards( )” which returns the evaluation for that state (if it is terminal).
[0028] The “isTerminal( )” expression in a report is true if the node is a leaf in the tree and requires evaluation, and “getRewards( )” which is given if the node returns an evaluation value (the expected reward) for the state if the state is terminal. While depth of node is used in the disclosure for illustrative clarity, other criteria may be used to determine the true value of “isTerminal( )”, such as detecting a projected end of conversation, session or communication. Given the four functions defined for a given domain, the MCTS algorithm will return the best initial action by simulating a “game” of carrying out actions and examining the rewards of the eventual outcomes. This process is currently the most successful method used by AI's that play games with complete information like Chess and Go and incomplete information like bridge, poker and Starcraft. MCTS automatically balances its search process between exploration of new move possibilities with exploitation of known successful paths. As a search method the family of MCTS algorithms are actively being researched and developed.
[0029] One insight of the current invention is that three of the four functions can make use of recent advances in broad coverage language models 110 to provide move generation 108 from contextual descriptions of an ongoing conversation or interaction and also to provide part of the necessary evaluation function required by the “getRewards” function of MCTS 106. In particular, by prompting the language model with the initial start (or prompt) of an evaluation phrase such as “I feel . . . ”, “I believe . . . ,” or any phrase that would normally lead to the statement of an ego-centric opinion on the state of the situation at that point in the dialog, an evaluative statement will be generated.
[0030] Another insight is that this linguistic/symbolic evaluative statement generated by the language model can be converted into a numeric value by application of a sentiment analysis (or other linguistic classification) function, which returns a numeric positive or negative value for a given statement. MCTS uses these evaluations of the leaves of the search tree to generate an expected average value for each possible initial action. Another feature of MCTS is that it can function as an “anytime” algorithm, which means the longer it runs, the better the estimated values, but at any time (e.g., due to time constraints) one can stop the process and receive plausible estimates for the best initial action. Process 120 extracts the highest ranked response from the MCTS search, along with other trace or log information. Process 120 corresponds to Step 004. This information is sent to the system response trace module 122 which will send the selected response to either the user or the self-play module 124, and to the conversational logs 116, for self-play-based learning.
[0031] Language model 110 corresponds to Language model 008 shown in FIG. 1 . The MCTS processor 106 fills the role of Step 003 shown in FIG. 1 . Optionally, a System Intent Processor (SIP) 130 may be included, to translate and interpret the utterances and control statements generated by the system into computational, retrieval or control actions. SIP processing may occur on either or both sides of the external user interface 101. The system intent processor is like giving the system its own Siri™ or Alexa™ type device. The SIP 130 turns the English output of the system (which may be a goal or direct command) into an action. The SIP 130 is a separate module that listens to the overall system output and executes requests it recognizes. Multiple methods exist for implementing such intent processors. The user interface 101 and the System Intent Processor 130 correspond to the Step 007 shown in FIG. 1 . The language model or models 110 used by 108 and 106 may be either statically fixed or continuously trained by a Language Model Training Processor (LMTP) 112. LMPT 112 can either accept training material from a corpus of reference conversations or examples of language usage 114, or from the conversational logs 116. The conversational logs 116 record the best system response given each input from the human 100, the self-play module 124 or other dynamic sources 118 such as internet search and extraction. The Language Model Training Processor 112 corresponds to the Step 011 which is shown in FIG. 1 . Storage 116 corresponds to Step 009 and storage 114 corresponds to Step 010. Information from the search analysis process in Step 120 can be used to create records to seed the self-play module 124 with start points stored in a start point data base/knowledge base (DB/KB) 126 or other stores for future exploratory conversations. The self-play module provides all the information required for the context packet 102 to continue exploration of possible continuations of conversations. An interface to the start point KB/DB 126 can be provided to insert new exportation points from external sources such as human users.
[0032] The use of LMTP 112 from source referencing conversations stored in data files 114 and conversational logs stored in data files 114 improve the operation of the language model 110 as both a move generator and evaluation generation model. In some uses of language model 110, the search process of 106 may be too expensive in terms of time. The LMTP process in fact causes the language model to lossily summarize, compress and compile the information derived from the tree search. As such, the trained language model 110 can be used for non-search applications on remote servers 128, where simple input/output operation of an enhanced language model is desired. Non-search applications for remote servers 128 may be realized through transfer learning, by training a simpler language model on the system output. The primary function for the system as a whole is analogous to the operation of the self-teaching game system AlphaZero/MuZero, except that instead of operating in a zero sum game space like Chess or Go, the space the described system operates within is the space of dialog, conversation and interaction (both virtual and embodied).
[0033] FIG. 3 is a flow chart depicting a process for generating a contextual description. Input information is accepted which describes the current situation and global context from various sources, in natural language form, background information, and a log of recent history, which may be empty when the process begins. Most language modeling processes developed to date either return an estimated probability for tokens by utilizing a flat probability or one which in some way is conditional upon a sequence of tokens over a window of some maximum depth. The goal of the contextual description generation process is to ensure tha tthe information most relevant to language model generation process is selected and fits within the processing window of the language model. The context generation process must rank all available information to ensure the most salient information to the current situation (and any problem posed) is presented in a package to the deliberation process and that it fits within any processing constraints of other modules.
[0034] The process begins with Step 200. In step 202 a summary is created of the most relevant information from recent history of the interaction stream. The summary captures the “gist” of the current situation. By way of a non-limiting example, in some implementations the Textrank text processing method maybe used to provide summarization of the collection of recent text being summarized. Other automated methods are known where a language model may be used to produce summarizations of text. Given that current State of the Art language models using the transformer neural model utilize an internal model of attention, it is useful for an attention-based summarization system to be employed. Textrank implements a simulation of where attention would be focused over a graph of nodes representing a corpus of text. Also, just as Pagerank has topic focused variants, Textrank can be modified to provide topic focused summarization. The primary benefit of using Textrank is the ability to provide extractive summarization over arbitrarily large historical logs.
[0035] The process then proceeds from Step 202 to Step 204 and begins to determine key word queries from the then current summary of the interaction stream determined in Step 202. The process proceeds from Step 204 to Steps 206-210. In Step 206 the current interaction stream is queried for relevant content to the queries. In Step 208 the history of a respective agent developing the conversation is queried for relevant content. Similarly, in Step 210 the background information database is queried for content relevant to the keywords determined in Step 204. The process then proceeds from Steps 206, 208 and 210 in parallel to Step 212 in which the results of the respective queries are merged. In Step 212 the query results are combined, ranked and sorted. By way of a non-limiting example, in some implementations TD-IDF based text retrieval methods may be used to query each background data source.
[0036] Then in Step 214 the results from the summary of the information stream from Step 202 is merged with the summary from Step 212 to provide current situational input from multiple relevant information sources. It is via this merge process in Steps 212 and 214 that information most directly relevant and associated with the current input from the recent conversations is made available to the context description. The merged output of Step 214 provides a final contextual description that fits the constraints of downstream processing elements. One such typical constraint is that the total length of the contextual description must be less than some maximum, usually determined by the processing capacity of language modeling technology or method. For example, current implementations of GPT-2 and GPT-3 have processing windows of generally less than 2048 tokens. The final merging process would select the highest-ranking set of collected information that fits within the 2048 tokens. A simple greedy sort and merge process is employed, though other methods such as evolutionary or other search methods may also be utilized, and is detailed later in FIG. 17 . The task of delivering the highest value product in a fixed sized carrier maps to the mathematical area known as packing problems in the field of combinatorial optimization. In particular we wish to maximize the relevant information presented to our language model for analysis and use the work done on the fractional knapsack problem; since while in the general case the optimization problem is NP-complete, the fractional variant admits a good polynomial time greedy approximations. Also, dynamic programming and genetic programming methods exist which may be of use for larger scale or different variants (possibly multi-modal problems), and may be relevant when the option set is continuously or asynchronously updated and an “anytime” update capability is desired.
[0037] FIG. 4 is a flow chart depicting an abstracted tree search process (of which MCTS is a variant) which utilizes the contextual descriptions in a deliberative search process to render and select a next action. or utterance for speech, based on the projected expected outcomes as determined by the search process. The text generation search process is provided by simulating an “internal dialog,” through a self-talk virtual imagination process. The process begins with Step 300. In Step 302 the contextual description of Step 214 of FIG. 3 and the root node of the search process are added to the process queue, along with a set of parameters for the search configuration. The search configuration will include a selection of one more language models on which the self-talk virtual conversation occurs. The search configuration would include everything necessary to define the initial search condition, the parameters of the search process, and the evaluation criteria and termination of search criteria. While defaults may be used, custom parameters may be specified for each and include information to define the initial node, any search constraints (the maximum resources to use in terms of time, memory, nodes to be processed, evaluation method to be used, expansion rate at each level, etc.), models used other than the defaults for generation and evaluation and information that might influence either (such as custom prompts), and response generation parameters (length of response generated, sampling parameters such as temperature, cutoff probability, etc.). The contextual description may include information regarding the selected personality and motivation of the virtual agent being simulated, as well as selected parameters and perceived parameters regarding the personality and motivation of the other various participants to the conversation exchange. For example, the virtual agent may be defined as being helpful, cheery and talkative. Such adjectives will influence the language model generation profile, which in turn influences the sentiment of text associated with the given character. Similarly, the projected and perceived traits of personality and motivation of the other various participants will also affect the simulated paths of the self-talk imagination process rendered by the search process. These assigned traits of the virtual agent and those of the participants may also be included in the search configuration. The initial parameters from Step 302 are input into Step 303 and the process queue is executed. The results of the execute process queue of Step 303 are then input into Step 304. Step 304 determines whether the termination criteria has been met; and if so, the process proceeds to Step 318 in which an options list is extracted from the current search results and in Step 320 the simulation will end.
[0038] If in Step 304 a determination is made whether the search criteria has been reached. The search termination criteria will that require at least one evaluation to have been performed and that some specified resource has been exhausted. Typically the exhausted resource may be real wall clock time expended, a number of node expansions, a lack of diversity in options (indicating total search space exploration), or convergence of the values assigned to the top level nodes. MCTS has the properties of being able to incorporate any of the above criteria into its termination condition. If a determination is made in Step 304 that the search criteria has not been reached, the process proceeds to Step 306, and the next node is selected for adding to the process queue. Next the process proceeds to Step 308, and a determination is made of whether the node requires evaluation. If a determination is made in Step 308, that the evaluation is not required for the selected node, then the process proceeds to Step 310 and a move generator is implemented to create one or more successor nodes. Then in Step 312 successor nodes are added to the process queue and the process proceeds to Step 303 to process the nodes in the process queue, and then after Step 303 the process returns to the search termination criteria evaluation in Step 304. If instead in Step 308 a determination is made that the node selected in Step 306 requires further evaluation, the process proceeds to Step 314 and a “generate evaluation method” is applied to the node. The process then proceeds to Step 316 and the values of the “generate evaluation method” are applied to the ancestor nodes of the node selected in Step 306. The process then returns to the process Step 303 and continues.
[0039] The search process uses a move generation method 310 to propose new actions that are then used to create new contextual descriptions by appending the proposed actions to the existing contextual description, changing the prompted actor variable as required. This describes a state space search, where the state is defined by an actual series of events and utterances and the simulated set of events and utterances are used to reach a certain point in time. The simulated events or utterances are generated by the language models predicting the logical continuation of the sequence used to reach a given state.
[0040] Each node in the state space search consists of a context description that contains the sequence used to reach the state represented by that node. Each node also contains ancestor node information where the ultimate ancestor node is the root node of the search tree and the root node defines the initial input to the search process. Each node contains information on the action that was taken to transition from the ancestor to the current node. This is the output of the language model that is appended to the ancestor node contents to generate the current node. Each node contains an evaluation value, indicating either the final evaluation of a leaf node in the search process or the estimated reward value of the descendent nodes. Each node contains a value signifying the distance which that node is located from the root node, along with visitation frequency information. This information allows depth constraints to be placed on the search process and the expected reward to be computed.
[0041] The move generation Step 310 is implemented and expands upon a node's context description and generates a new node with a new context description by using the node generation process and appending the output generated by the language model, when the distance of the expanded node from the root is less than some threshold or when some other node expansion criteria is met. Changing the prompt used to generate a node expansion may be fixed (implementing say verbal taking turns) or left at the discretion of the language model.
[0042] The search process may be implemented with the described node data structure and a queue to hold lists of nodes to be processed. The simulation Steps continue until some termination criteria is met in Step 304, such as expiration of time or exhausting some resource. If the termination condition does not hold then a node is selected from the processing queue for expansion Step 306. The content of the node is examined, and based on depth (or other criteria), a determination of if the node qualifies as a leaf node and requires evaluation or is an interior node and requires expansion. If the node requires expansion, then the afore described node generator step Step 310 is applied to generate one or more new nodes, each with parent/ancestor being the current node that was being examined. Each of the new nodes is inserted into the process queue 312 and control is returned to Step 304. If, however, Step 308 determines that the node is a leaf and requires evaluation, then a separate “evaluation prompt” is used in Step 314. The evaluative prompt is a prompt to the language model to summarize the situation described by the contextual description of a node in a way that is easy to evaluate using either sentiment analysis or other language classification methods. Once evaluations are made in Step 314 and Step 316, values are propagated back through each of the linked ancestor nodes to update each ancestor node value.
[0043] Once the search termination criteria is reached, the search process returns in Step 318 an option list consisting of the language model derived actions that lead to the immediate descendant nodes of the root node, along with each of their expected long-term values as estimated by the simulation process of each action. Ultimately the agent will use the option list generated to select the next action either deterministically or probabilistically. By way of a non-limiting example, the afore mentioned search process in some implementations may use Monte Carlo Tree Search to implement the optimal action search process in FIG. 4 .
[0044] With regard to Step 306, it is known that various criteria used to select the next node to process produces various search behaviors (depth-first search, breath-first search, best-first search, A* search, MCTS search). The data structure of the node is modified to contain such data as to make the desired search behavior possible, notably (but not necessarily limited to) the nodes distance from the root, estimates of the average value of child nodes, and estimates of the value for leaf nodes. Optionally, any heuristic value of distance from some goal state may be included; however, the estimate of average descendant value may be used as a proxy.
[0045] With regard to Step 316, different propagation behavior results in the emergent property of different overt behaviors. In typical planning for two interacting agents, each agent may choose to minimize or maximize its expected return. This determines the style of conversational game being played and behavior observed. The following “TABLE I” lists by way of example, Evaluation Propagation Policies and a corresponding Behavioral Descriptions:
[0046]
| TABLE 1 | ||
|---|---|---|
| Evaluation Propagation Policy | Behavioral Description | |
| Maximizing the sentiment of all actors | Win-Win Optimist | |
| Maximizing the sentiment of the user | Comedic/Entertaining/ | |
| (while ignoring itself) | Subservient | |
| Maximizing the sentiment of system | Narcissistic/Self-centered | |
| outputs (while ignoring the user) | ||
| Maximizing system sentiment and | Sadist | |
| minimizing user sentiment | ||
| Maximizing user sentiment and | Masochist | |
| minimizing system sentiment | ||
[0047] The evaluative prompt used in Step 314 need not be limited to just the one actor prompt but can be prefixed with material to influence the evaluation given. This ranges from suggestions to select from a range of evaluation options (“How was the experience: Good, Neutral, Bad? Daxine:′ I believe the experience was [EOP]”), to prompts to generate a more open ended scene description (“As the scene closed [EOP]”).
[0048] FIG. 5 is a schematic diagram depicting different layers of an interchange MCTS processor 338. The MCTS processor 338 conducts a search based on the description of the individual state, the move generator and the evaluation function. The process of generating the initial description used in the root is performed by module 340. The MCTS processor 338 generates multiple levels N, 1 through Level 0 of the search tree to determine the best response to input. For illustration it is assumed that the MCTS processor 338 is instructed to generate searches of depth N from the initial root node. A typical MCTS may be used without a fixed search horizon. The MCTS processor 338 generates the first layer of Level N−1, and the best response extractor process 342 returns the highest-ranking node, or some selection based on the relative ranking of the nodes in level N−1. As in the normal MCTS process, a number of nodes are generated for one or more additional levels labeled as levels 344 using the language model to create the search tree out to some search horizon, which has the terminal level 348 (“Level 0”). At the pre-terminal level 346 (“Level 1”), the evaluation prompt is used for the nodes of 346 to generate the evaluation terminal nodes of 348. The generated nodes of level 348 (“Level 0”) thus contain the text or other information which is converted to a numeric value through a symbolic to numeric conversion process. In the illustration that conversion is done using a sentiment analysis process. For instance, assuming the system is simulating a character named Bob, the evaluation prompt might be:
-
- [0049] Bob: “I feel[EOP]”
where “[EOP]” represents the end-of-prompt and is where the language model would insert its continuation.
- [0049] Bob: “I feel[EOP]”
[0050] The Language Model (LM) is prompted to complete the evaluation sentence, based on the total context provided for that node. The text values are then converted to a positive or negative numeric value by the sentiment analysis process, (or any other leaf evaluation process) and the value is propagated back up the tree to update the level N−1 node values.
[0051] FIG. 6 is a flow chart depicting a move generation process using language models and describes a method of querying one or more language models or knowledge sources, such as using a search tree process, such as with an MCTS processor, to provide the equivalent of “move generation” or node expansion required by the search process. The move generation process begins in Step 400. The language model or knowledge source is presented with a contextual description and is prompted to generate a continuation as output. In Step 402 a prompt is presented which identifies the agent by name or role, (i.e. “John: ‘”, or “Doctor: ’”, etc.) and is in a format matching the actor identified responses in the language model training data 404. In Step 406 the language generation prompt is preformed and formatted and input to Step 408. In Step 408 the output may optionally be truncated if a change in actor is detected in the generated output, such that the output of the entire process contains only one valid conversational or interaction “move”.
[0052] Optionally, the response may be generated by rule-based, retrieval-based or other sequence-based generation methods (e.g. chatbot technology, question-answer engine, etc.). The operation of language models can be substituted with rule-based, retrieval based and language model-based continuation or response generation methods. The initial set of options at the root can also be generated by other means, including but not limited to rule-based, retrieval-based or language model based. Also, a fixed set of options may be included or appended to the existing list, such as “unknown” or recognized sets of standard operations (e.g. via user intent detection methods).
[0053] FIG. 7 is a flow chart depicting an evaluation process using language models and outlines the Steps required to query a language model or knowledge source for a situational evaluation as required in Step 314 of FIG. 4 . FIG. 7 describes a method of querying one or more language models or knowledge sources to provide the equivalent of an “evaluation” function. The situational evaluation process begins in Step 500 and then proceeds to Step 502. In Step 502 the language model is presented with a contextual description and is then prompted in such a way that the language model will generate an evaluative output to Step 504. The prompt identifies the evaluative agent by name or role in Step 502 and is in a format matching actor identified responses in the content the language model has been trained upon. In Step 504 the name of virtual agent for evaluation is included. In Step 506 the prompt is formatted to match the identified responses of the simulated actor or agent in the language model used for training. In Steps 508 initiation of an evaluative statement or language use is included. In Step 510 returned text is analyzed to generate one or more values representing evaluation of the contextual description. An example might be “John: ‘I feel [EOP]”, where “[EOP]” represents the end-of-prompt. In such a case the language model would generate a continuation such as “satisfied with the outcome” or “sad about what happened”.
[0054] The output of such a generation Step is linguistic or abstract in nature and must be converted into a comparative value. The text returned is analyzed to generate one of more numeric values representing the relative evaluation of either the final evaluative statement or the entire contextual description thus generated. The evaluation process in the initial implementation utilizes sentiment analysis applied to the evaluative statement. Subsequent versions utilize sentiment analysis to train a backoff language classifier since, during processing, a large volume of content is generated and processed. However, the rule-based sentiment analysis system may not return a score in all cases. Examples of text that generate high sentiments scores with high confidence are used to train in an unsupervised way a backoff language classifier. The backoff classifier is used to provide low value estimates when the rule-based system can provide none. However, the language model-based classifier may be used as a generic programmable substitute for the evaluation process and may be trained on both the output of sentiment analysis or statement certainty and modality.
[0055] FIG. 8 is a flow chart depicting steps for evaluating the relative evaluation of the contextual description and expands on Step 510 in FIG. 7 . The analysis begins in Step 600 and then proceeds to Step 602. In Step 602 the text is analyzed using basic sentiment analysis and then proceeds to Step 604. In Step 604 the text is analyzed using goal oriented analysis and then the process proceeds to Step 606. In Step 608 sentiment analysis of Step 602 and the goal oriented analysis of Step 604 are merged together into a weighted evaluation.
[0056] FIG. 9 is a flow chart depicting evaluation utilizing a basic sentiment analysis which expands upon on Step 62 of FIG. 8 . The process begins with Step 700 to perform a simple rule based sentiment analysis upon a segment of text to indicate the total polarity of the text ranging from −1 to +1. The text is input to Step 702 which then tokenizes the text into a list of words and recognized phrases, creating a list of tokens. The lexicon contains a valence value (−1 to +1) for each word in the recognized vocabulary. In Step 704 the valence value of each token is looked up in a valence lexicon or a valence table. The valence values are then input to Step 706 in which the overall sum of the valences is computed as the sum of the valence of each token in the sentence. The process then proceeds to Step 708 and a normalized score is computed for the text being analyzed using the following equation:
NORMALIZED_SCORE=SUMS/sqrt(SUMS*SUMS)+ALPHA
where ALPHA=15 and is the approximate max value expected. In Step 710, the NORMALIZED_SCORE is clipped, so as to remain in the range (−1, +1). In 712 the NORMALIZED_SCORE is returned. The results of the Steps of 700-710 result in the general positivity and negativity being returned for a given statement expressed in text, relative to the values placed in the valence lexicon. By way of a non-limiting example, in some implementations a rule-based analysis like a VADER analysis, or another sentiment analysis method may be utilized.
NORMALIZED_SCORE=SUMS/sqrt(SUMS*SUMS)+ALPHA
where ALPHA=15 and is the approximate max value expected. In Step 710, the NORMALIZED_SCORE is clipped, so as to remain in the range (−1, +1). In 712 the NORMALIZED_SCORE is returned. The results of the Steps of 700-710 result in the general positivity and negativity being returned for a given statement expressed in text, relative to the values placed in the valence lexicon. By way of a non-limiting example, in some implementations a rule-based analysis like a VADER analysis, or another sentiment analysis method may be utilized.
[0057] FIG. 10 is a flow chart depicting evaluation utilizing a goal oriented analysis and expands on Step 604 of FIG. 8 . The analysis details various methods to analyze the returned text to generate one or more-values representing relative evaluation scores of the contextual description based on a basic goal-oriented analysis. In such an analysis the purpose is to return a value that indicates relative distance from some goal state (factuality, agreement, statement of ground truth, etc.). Various analysis methods may be employed individually and independently or in conjunction and in parallel. The process begins in Step 800 and then proceeds to Step 802. In step 802 the process receives a textual description to be evaluated and is passed to one or more subprocesses that perform text comparison. The process will then proceed in three parallel paths to Step 804, Step 810 and Step 814, respectively. In a first parallel path beginning with Step 804 the text is converted into a vector representation which is passed in parallel to Step 806 and Step 808. In Step 806 the distance between that vector and a relatively small set of reference vectors, and the nearest distance is returned. Here, vector refers to vectors encoding semantic information and numerous methods are known to generate them (word2vec, sent2vec, doc2vec, GLoVE, word vector embeddings used by various language models, random projections, etc.) A small reference set may be derived from a small set of initial target statements input at the start of the processing. The output of Step 806 is output to Step 818. In Step 808 the vector is used as a query to a larger vector database, which may represent longer-term goals or evaluations, and then output to Step 818. By way of a non-limiting example, in some implementations Approximate Nearest Neighbor search methods may be used in determining the most relevant vector-based match from a database or reference set. The second parallel path begins from Step 802 and moves to Step 810, and in Step 810 text based matches are found in a goal reference database. The matches found in Step 810 are input to Step 812 in which the matches are selected or merged, and then output to Step 818. Text-based queries may be applied to a database of reference sets of the form (text→value) or (vector→value). The output of the database query is ranked and merged into an output set. By way of a non-limiting example, in some implementations a language model may be trained to receive a goal statement and a textual description to be tested and that evaluative model would return an evaluative response for the value of the text test relative to the goal text with an interpretation process similar to that for sentiment evaluation employed. (i.e. v=sentiment (eval_model (goal statement, input))).
[0058] In the third parallel path moving from Step 802 to Step 814. In Step 814 a goal evaluation trained language model is used to find text based matches which are input into Step 816. In Step 816 the matches from Step 814 are converted using sentiment analysis methods. In Step 814, a traditional text classification system is employed where such a system utilized simple features, such as character and word n-grams and simple linguistic features. Text classification methods based on naïve bayes, perceptron, support vector machines, k-nearest neighbor, TF-IDF and others may be employed. In Step 818 the output of the various analysis methods are merged into a final weighted evaluation score and returned in Step 820.
[0059] FIG. 11 is a flow chart which depicts merging multiple analyses into a weighted evaluation and expands on Step 818 of FIG. 10 by detailing a simple method for converting multiple candidate results from multiple sources into a single weighted output value base on the similarity or distance metric provided by each. The process begins with Step 900 and the proceeds to Step 902 in which the list of matches with values and metrics are input. The metrics used are preferably either distance or similarity based. In Step 904 a determination is made of whether the metric is distance based or similarity based. If in Step 904 it is determined that the matched condition is distance based, the process proceeds from Step 904 to Step 906. In Step 906 the distance based metrics are converted into similarity based values in Step 906 and Step 908. In Step 906 the maximum distance for a candidate match is determined and input into Step 908. In Step 908 for every candidate a similarity value is determined. The two determinations for Steps 906 and 908 are shown in the equations below:
[0060]
The similarity values of Step 904 and the similarity values determined in Step 906 and Step 908 are input into Step 910, and an average similarity value is determined. In Step 912 a weighted similarity value is determined for each match, and then in Step 914 a combined weighted estimated similarity value is computed.
| For Step 906: | maxDistance=max  i  distance[i]  | |
| For Step 908: | similarity[i]=maxDistance-distance[i] | |
The similarity values of Step 904 and the similarity values determined in Step 906 and Step 908 are input into Step 910, and an average similarity value is determined. In Step 912 a weighted similarity value is determined for each match, and then in Step 914 a combined weighted estimated similarity value is computed.
[0061] Given K value estimates in value[ ] and the positive value array similarity [ ] the following are computations performed in Steps 910, 912 and 914:
[0062]
In Step 916 the combined, or merged, weighted similarity value is returned to Step 8l8 ofFIG. 10 .
In Step 916 the combined, or merged, weighted similarity value is returned to Step 8l8 of
[0063] FIG. 12 is a schematic diagram depicting an apparatus for self-learning through application of a self-play feedback loop. The “Replay Buffer” 900 is the memory of conversation logs 902 (corresponding to conversation logs 116 FIG. 2 ) and special training logs provided by reference conversations 904 (corresponding to conversation logs 114 in FIG. 2 ). The replay buffer 900 preferably includes both the human sourced corpora and the self-play logs. The “Shared Storage” 908 is the set of generated models 910 (corresponding to the language model 110 of FIG. 2 ). It contains any fine-tuned language model networks trained on the sum total of the Replay Buffer 900. One or more Self-Play engines 912 (corresponding to self-play engine 124 in FIG. 2 ), takes the latest language model image 910 from “Shared Storage” 908 and generates new conversation logs 902. It may do this using a special “interesting prompt list” provided by the start point data 126 of FIG. 2 , where different prompts are used, or it can use unconditional prompting. The output of each self-play engine run goes to its own section in the Replay Buffer 900.
[0064] The Trainer process 906 (corresponding to the language model training process 112 in FIG. 2 ) does a scan of the text in the Replay Buffer to create a training corpus, trains the language model training process 112 of FIG. 2 (Transformer-based, probabilistic, RNN or otherwise), and generates new language models 910 which are model images kept in Shared Storage 908. In addition to the internal self-play engines 912 additional External Interaction Engines 914 may be included which encapsulate systems that use the language models 910 to interact with the external world, and generate additional conversational logs 902. There may be one or more self-pay engines 912 and External Interaction Engines 914 existing and running in parallel in a given system.
[0065] FIG. 13 is a schematic diagram depicting an apparatus for interlocking training and self-play processes. A trainer process loop 1000 loads its initial parameters on startup 1010 and generates a reference corpus from the relevant Replay buffer set 1012, then trains the language model for N Steps 1014, writes the resulting updated language model out 1016 as a new reference model in the shared image file 1006. A determination 1018 is made as to whether a new corpus for training is generated 1012 or the existing corpus is used to continue to update the existing model 1014. The training process 1014 which is stated as being for N steps, can be any of the standard criteria such as N-steps, T amount of wall clock time, reduction in measure error past a threshold, etc. A number of restart criteria 1018 can be specified. One criteria could be based on a fixed number of iterations. Another criteria would be on the amount of new content added to the Replay Buffer 1004. It may also restart the process after every T-time units. Also Step 1014 may terminate early based on time or change in the replay buffer contents.
[0066] The Self-play engines use a similar processing loop 1002. First, they load their initial parameters 1020 and then load the latest shared image 1022 provided by the Training Process 1000. Each Self-play engine simulates a number of interactions 1024, recording each as a trace file stored in the Replay Buffer 1026,1004. Each replay engine then determines if the model it is using is “stale” relative to available models 1028 and, if so, loads the new language model 1022; otherwise, it continues to generate new examples using the existing model 1024. By having the Trainer and Self-Play engines implemented as independent processes provides a number of benefits. The interoperation of the two roles are “near embarrassingly parallel”, since other than the startup either role can continue to generate useful work without requiring blocking synchronization from the other. They can provide the option of running continuously and asynchronously. They can operate on a single system or run on multiple independent networked processors or computers either centrally located or widely distributed. Such an arrangement provides the benefits of natural flexibility, scalability and redundancy.
[0067] Fortunately, the self-play process requires less processing power than training to generate content and its task can be performed by an individual processor much more modest than Trainer. The Trainer on the other hand benefits from being a central process with maximal processing power or efficiency (a GPU/TPU system). The pool of self-play engines can be implemented locally on one system or on a cluster of multiple computers of different processing power. A cluster of self-play engines may be either tightly coupled or loosely coupled and widely distributed.
[0068] Ideally, it is desirable to have a simple, easy to configure system where one can simply add self-play engines to a virtual network and they start producing input to the Replay Buffer 1004. This can also be implemented as a Grid computing model, where during initialization 1020 and model testing 1028 each remote self-play engine would retrieve a URL for the current best available language model. It starts producing output, compresses and uploads the result to an Replay Buffer implemented through an internet cloud based repository. CloudTurbine provides an example of implementation of a distributed system using a cloud-based file-sharing communication model, (such as Dropbox or Amazon S3). Dedicated network file-storage infrastructure of course may also used.
[0069] Another modification is to have individual Trainer processes sub-sample the entire Replay Buffer, so the training set it uses is representative but does not require the use of the whole buffer. This can allow faster cycling of the training loop 1000 since the corpus construction phase can be substantial as time passes and the Replay Buffer 1004 increases in size. One can also have a sliding window or exponential decay weighting for the sampling process, such that older data is given lesser weight than new data. Such a system would also allow a core set of data in the Replay Buffer to be given a fixed constant priority. Such a preference weighting could be used to give direct experience from external sources 914 a preferential weighting over self-play 912 generated traces.
[0070] FIG. 14 is a schematic diagram depicting a design of an agent for generating linguistic responses to verbal and situational input and illustrates the use of the deliberative process with a simulated agent. User x01 corresponds to user 100 in FIG. 1 . Simulated agent x02 may be physically embodied, virtually simulated in virtual reality or augmented reality, or wholly exist in an abstract simulated state. User x01 provides inputs to simulated agent x02, which x02 captures and encodes via processes in x03, to create the current situational input represented by x04. X04 provides information to x05 to select information to form the context description from relevant background knowledge and recent history. Process x04, x05 and x06 corresponds to Step 002 in FIG. 1 , and Steps 102 and 104 in FIG. 2 . X07 is the deliberative search process and corresponds to Step 003 and processor 106. X08 is the action extraction and selection process and corresponds to Step 005. X09 corresponds to the short- and long-term experience memories. X10 translates the action selected by x08, and executes them either as verbal actions, animations or virtual operations (i.e. web search, directed calculation, control operations, etc.). X10 corresponds to 007. The primary purpose of the system is to allow the simulated agent x02 to exhibit more intelligent behavior through simulation and evaluation of the near-term consequences, through the use of a situational descriptions (optionally encoded as natural language) using collected experience and knowledge (especially collected and encoded in natural language).
[0071] FIG. 15 is a flow chart depicting a training process required for a broad coverage agent using predictive language models. Some embodiments utilize one or more trained generative or predictive language models which are trained on one or more data sources. Such data sources need to include examples of producing next-Steps actions, responses and evaluative statements in the target domain (which may be very broad). As such the corpus of training material will determine the performance of the overall system, both by its selection of initial options and actions, the plausibility of the continuations used in node expansion, and the eventual evaluation statements prompted at the leaves of the search process. It is thus recommended that for broad coverage applications (general chatbot, personal companion, etc.) that Steps y02, y04, y06, and y08 be employed.
[0072] Optionally, y10 may be employed where the system operates in an internal feed back loop to generate simulated conversation logs which are used as additional training material. Self-play provides additional material to reinforce the initial option generation function to produce output more likely to lead to positive outcomes. Unsupervised self-play combined with state space search has shown excellent results in games requiring human interaction.
[0073] FIG. 16 is a schematic diagram for providing an informational funnel to construct Immediate Situational Descriptions and details the informational funnel, from external sensors to immediate situational description. Simple modifications of the context description provided to a language model can produce very different results. During normal operation in chatbot mode the interaction is presented to variants of OpenAI's GPT system in what we refer to as “screenplay format”, consisting of lines with the speaker's name and what they said. A normal input prompt contains the back and forth conversation log followed by the bot name and an open quotation mark, asking the system to generate what the bot is expected to say. The conversational log format is what is collected from users and what is collected from both internet and textual sources. Part of the training dataset for a general chatbot includes movie and tv transcripts. However, we also include normal text, fiction (from various sources, lengths and formats), and a corpus of novels from various genres of literature (action, romance, sci-fi, adult, etc.).
[0074] While conversations with the system has good quality, when employed with an embodied agent the output generated lack a certain situational responsiveness. This was to be expected, given the information of only the words spoken between actors. A slot for an immediate situational description was placed between the human actor spoken line and the prompt for bot's response. This situational description corresponds to what would be immediately sensed by an embodied agent, such as “She could see him looking at her”, “She could feel his touch” or “She was in a happy mood.” Such situational descriptions added between the actor's lines significantly improved the output and responsiveness, even though the system was not trained on text in that format. The additional input helps the system interpret the prior input and conditionalizes the generated output.
[0075] Having this slot provides a way for placing non-verbalized sensing and motivational information to influence response generation. In the embodied agent use case, this additional information slot provides for the inclusion of simple mappings from sensor input and user descriptions. The flow outlined below is similar to Baar's Global Workspace attentional funnel.
[0076] Z00 corresponds to the sensory input of the simulated agent x02, and contains inputs from different sense modalities (audio, visual, tactile, kinematics, etc.). In addition, z02 contains input from various non-symbolic emotional simulations while z04 contains input from various motivational simulations. Information from these sources may be collected in z08 (which corresponds to x03), which acts as a sensory buffer, and encapsulates both the verbal and non-verbal description of the systems state. In addition, other language processors (z06) may be in employed and operate in parallel. The various sensory inputs in z08 are translated into verbal equivalents and merged with any inputs from z06 to form a complete list of linguistic descriptions and translations that describe the immediate situation at some time T (z10). This list is priority sorted and merged into a final Immediate Situational Description (ISD) at z12 which forms the input to x04 and part of the scene description of step 102 in FIG. 1 .
[0077] Normally the interaction log presented to the language model is in screenplay format. The ISD is inserted between the actors lines to provide necessary contextual annotations to include information that is not captured by the normal dialog exchange.
[0078] FIG. 17 is a block diagram for generating and selecting content using entropy density and a fractional knapsack model. A key requirement for utilizing language models with fixed window processing constraints is selecting the most important or relevant information elements for the system to process. This task can be mapped into a modified fractional knapsack problem where the goal is to maximize the information contained in a fixed size processing window. This is an integral concept for 002, 204, and z10-z12. The selection process w00 accepts a list of strings and a total capacity in w02. For each string in the list of strings accepted, it creates an element that contains the length, entropy, entropy density and string index for future reference. In information theory, entropy represents the “surprise” or information content of some event E. This is usually represented as the negative log of the probability of an event. In this case the probability is of a word's occurrence relative to some natural language corpus. To map into a fractional knapsack formulation, the concept of entropy density is created, where the average entropy of each token in each string is estimated. This allows the information content of each string to be compared and the greedy algorithm to be applied in w06 and w08. It is noted that w08 varies from the standard algorithm to admit whole strings. One can admit the standard algorithm if one is allowed to simplify, summarize or otherwise modify the final string selected by the standard algorithm in such a way that it meets the final fractional assignment requirements.
[0079] Once w08 completes, the results and final value are returned in w10. The list returned are indexes of the original string list passed that should be included in the processing window. The sorting operation of w06 may be modified to either sort in ascending or descending order. If ascending order is used then the lowest entropy strings are selected first and the most probable events are attended to. Likewise, if the descending order is used, then the most improbable and individually informative events are included. This allows the effects of attentional low and high pass filtering to be realized through symbolic means.
[0080] Also, the sources of information accepted by w02 may be heterogeneous and only require the ability to estimate the probability of each element under consideration (in the example, text). The method may be applied to text from multiple languages and events of multiple modalities that can have their entropies measured. The probability may be estimated via uni-gram, bi-gram, n-gram, other language models or any method that can return a probability for each string and thus estimate the entropy. Metrics other than entropy density may also be used. Also, the knapsack planner can be applied recursively over subsets of relevant information. For instance, a certain capacity may be reserved for certain information types in a global context description. Given a list of pending goals or unknown variables, the relative value of the acquiring each may be specified and the most important combination presented, each using a portion of its own reserved capacity. Any capacity that may be unused may be passed on as extra capacity for other slots in the global context description.
[0081] Msc Notes Re Knapsack
[0082] Primary problem addressed:
-
- [0083] for chatbot applications language models generate inappropriate responses
- [0084] for chatbot applications language models being insensitive to the consequences of their output
- [0085] for embodied agents (real/virtual) having language models responding to sensory input, emotional and motivational simulations
- [0086] for language models to control various aspects of embodied agents and associated systems
- [0087] Dynamic real-time unsupervised training of evaluation classifier
- [0088] External specification of evaluator by providing training of a evaluation classifier
- [0089] Using the trained classifier as a backup to the primary classifier and to provide gradient information
- [0090] Using language models for summary generation
- [0091] Match of attention between textrank and gpt/transformers
- [0092] Translation of sensing into NL and immediate description
- [0093] Dynamic specification of the formula used in the evaluation process
[0094] Information packet between user facing front-end and backend specifying [classifiers used, eval function, profiles, roles, etc.]
[0095] Specification of prompt and expansion factor at each tree level
[0096] Definition of language model to include transformers, rnn, markov models, rule-based, info retreieval and any means of generating a response sequence to an input sequence.
[0097] Using the principles of evaluation but with single threads, i.e. NON-tree search but N-single extensions with evaluation applied to the end of each
[0098] Use of GA process in selecting best subset of text for the processing window. In fact we can see it as a Fractional knapsack problem where:
-
- [0099] weight=length of text,
- [0100] capacity=window size
- [0101] value=priority/info content
[0102] In our description one can view a language model to be a blackbox that generates a response sequence to an input sequence. There are many methods that can be used to fill that role, using n-grams, rules and patterns, probability networks (markov models), information retrieval systems (TF-IDF,LSA), and neural nets (rnn's, LSTM's, transformers). While implemented with large scale trained generative pre-trained models like GPT, the architecture still can be implemented with any of the other methods.
[0103] In normal operation, the GPU system can generate a long continuation for each prompt. The test system normally generates batches of 4 continuations at a time until it produces enough to meet the limit specified for a given node (say generate 16 possibilities). Each time it produces a continuation, it generates a little bit more (say it is told to generate 64 words), and the extra is trimmed down to make one complete response. This makes the system behave as required for normal tree search.
[0104] However, one can request the system produce enough content to reach the normal leaf node (say 256 words). In this case you can break the content produced into turns and apply the evaluation to the content of the leaf. This is the equivalent of looking at just one possible future for each option, versus looking at a branching tree of futures for each option. The benefit of course is each N options require N generation Steps, instead of N*level1*level2* . . . * level_max generation Steps. This allows the system to be more responsive without having to give up the future projection ability.
[0105] One can also truncate the generation of content in the middle levels, as long as the content generated is sufficient to “sketch the outline of future development”. In some cases it is sufficient to just detect the fact that an argument will ensue, not the full details of the argument. By specifying a truncated generation process, one can speed up the overall search response.
[0106] Through the use of iterated search one can use smaller/simpler models. The system internal dialog simulation allows it to predict negative outcomes and give preference to responses that lead to positive outcomes.
[0107] Information passed from the front end to the processing system to generate a response may include:
-
- [0108] Maximum depth and expansion at each level specified
- [0109] Specify the evaluation function by selection or by formula to be interpreted
- [0110] Prompt to be used at each level
- [0111] Literal prompt
- [0112] Default back and forth prompt
- [0113] Use of rule based, information retrieval or 1m based generation
- [0114] Option to let a rule based system evaluate the next generation Step
- [0115] Additional sensory information or summary of sensory information in the form of NL text
- [0116] Interpretation of NL text to generate actions including animation
[0117] In a practical application the linguistic actions may be specified, along with linguistic descriptions of their outcome:
-
- [0118] User: It is dark in here.
- [0119] System: House, turn living room light to maximum.
- [0120] Living Room light is at maximum. The Living room is very bright.
- [0121] User: Ah, much better.
[0122] The goal-oriented evaluation processor may be trained on material annotated from sources to indicate the achievement of some goal in some interactive session. For instance, achieving some outcome a user may desire, such a solving some problem or performing some service. For instance, achieving some customer service objective (problem resolution, sales, persuasion, etc.) Labeling of such material may be achieved through explicit means (direct confirmation of resolution) or implicit means (user not complaining about problem over some time period). In situations with long or continuous interaction histories, persistent changes in user sentiment may detected and used as a reinforcement signal.
[0123] The present invention provides advantages of a simulated virtual agent which responds with textual expression to textual input from external user and external virtual agents in combination with the audio, visual, and tactile inputs. The textual expression from the virtual agent being generated by processes utilizing one or more language models, self-moving modules, and both current and prior conversation logs with self-learning reference conversations in combination with MCTS processes and self-learning techniques. Selection of the virtual agent's textual responses including simulated sentiment analysis and goal oriented analysis to assign values which are weighted, normalized, and merged in selecting textual responses through internal self-discussion tree search techniques.
[0124] Although the preferred embodiment has been described in detail, it should be understood that various changes, substitutions and alterations can be made therein without departing from the spirit and scope of the invention as defined by the appended claims.
Claims
What is claimed is:
1. An Artificial Intelligence (AI) language virtual agent having self-improvement features which incorporate language modeling and tree search techniques, the AI language virtual agent comprising:
a data processing system having program instructions which provide the AI language virtual agent;
sensory inputs having one or more audio, visual, and tactile receptors located proximate to the virtual agent for collecting indicia of verbal and situational indexes, wherein said sensory inputs are connected to said data processing system for inputting the indicia of verbal and situational indexes;
wherein said data processing system is adapted for processing said indicia of verbal and situational indexes to define a current situational description;
said current situational description including natural language input, qualities of the virtual agent, and indicia regarding subject matter context of a present conversation, wherein said qualities of the virtual agent include temperament and textual tendencies, and the indicia regarding subject matter context includes textual logs from recent conversational exchanges;
a database adapted for access with said data processing system having one or more language models, conversation logs storing text from prior textual exchanges, and reference conversations for training;
one or more self-play engines for training said one or more language models with self-play;
one or more external interaction engines for communicating with one or more external agents;
one or more self-moving modules for advancing the method of external agents communicating with the virtual agent via a combination of textual exchanges; and
wherein the AI virtual agent utilizes tree search techniques in combination with said one or more language models for outputting textual responses to the current situation description.
2. The AI language virtual agent according to claim 1,
wherein the virtual agent utilizes Monte Carlo Tree Search (MCTS) techniques in combination with the one or more language models for outputting textual responses to the current situation description; wherein the virtual agent responds with textual expression to verbal input in combination with the audio, visual, tactile, and other sensory inputs after merging an analysis of basic sentiment analysis and goal-oriented analysis of the AI virtual agent; and
wherein the sentiment analysis is determined by the text being tokenized into a list of words and phrases, a sentiment value is assigned for each token in which a summation of the sentiment values are computed and then normalized as compared to an expected value to determine respective normalized scores, and then the respective normalized scores are compared to one or more indexes which are summed to provide discrete output values which provide a respective sentiment value for the text being analyzed.
3. The AI language virtual agent according to claim 1, further including the steps of:
selecting an optimal response to a current situation description with respect to immediate and non-immediate goals or outcomes by using one or more language models to form the optimal response search process;
wherein the language models are used to anticipate future outcomes and future responses;
wherein the language models are also used to evaluate the desirability of projected future outcomes;
converting the language models evaluation into a format suitable to guide tree search methods; and
wherein information regarding the desired goals or outcomes may be included in the current situational description.
4. The AI language virtual agent according to claim 1, further comprising the steps of enabling the evaluation of possible responses, for use in a real or virtual agent which include the use of state space based search evaluation methods.
5. The AI language virtual agent according to claim 1, wherein the state space based search evaluation methods include MCTS processes.
6. The AI language virtual agent according to claim 1, further including the steps of:
constructing the contextual descriptions to maximally represent the current situation while meeting any processing constraints of the language models used;
providing nodes of the states of the search space which contain contextual descriptions and search control information;
receiving an initial starting contextual description and search control information; and
exploring the state of possible future contexts by using one or more language models provided with sufficient information in the form of the contextual descriptions and prompts that cause the generation of new situation descriptions.
7. The AI language virtual agent according to claim 1, further comprising:
utilizing one or more language models to generate evaluative statements or annotations on the desirability of a situational description;
translating symbolic evaluative statements into numeric or vector format suitable for use in common state space search systems;
selecting an initial action from an initial set based on the outcome of the complete search process;
performing an action or set of actions based on the initial action selected;
recording the system and environmental interactions to one or more logs;
providing for the continuing training of the language models used by the system using reference material and direct experience; and
providing self-improvement of the entire process through the iterative retraining of the language models, through the use of self-play to explore the space of novel possible situations and to discover paths towards positive resolutions in those situations.
8. The AI language virtual agent according to claim 1, wherein said virtual agent responds to verbal input in combination with various sensory inputs, including one or more of audio, visual, tactile inputs, with said combination expressed in natural language and merged with background information to create a contextual description; and
wherein in addition to textual expression the virtual language agent responds with animation and virtual effectors.
9. An Artificial Intelligence (AI) language virtual agent having self-improvement features which incorporate language modeling and tree search techniques, the AI language virtual agent comprising:
a data processing system having program instructions which provide the AI language virtual agent;
sensory inputs having one or more audio, visual, and tactile receptors located proximate to the virtual agent for collecting indicia of verbal and situational indexes, wherein said sensory inputs are connected to said data processing system for inputting the indicia of verbal and situational indexes;
wherein said data processing system is adapted for processing said indicia of verbal and situational indexes to define a current situational description;
said current situational description including natural language input, qualities of the virtual agent, and indicia regarding subject matter context of a present conversation, wherein said qualities of the virtual agent include temperament and textual tendencies, and the indicia regarding subject matter context includes textual logs from recent conversational exchanges;
a database adapted for access with said data processing system having one or more language models, conversation logs storing text from prior textual exchanges, and reference conversations for training;
one or more self-play engines for training said one or more language models with self-play;
one or more external interaction engines for communicating with one or more external agents;
one or more self-moving modules for advancing the method of external agents communicating with the virtual agent via a combination of textual exchanges; and
wherein the AI virtual agent utilizes Monte Carlo Tree Search (MCTS) techniques in combination with said one or more language models for outputting textual responses to the current situation description.
10. The AI language virtual agent according to claim 9, further including the steps of:
selecting an optimal response to a current situation description with respect to immediate and non-immediate goals and outcomes by using one or more language models to form the optimal response search process;
wherein the language models are used to anticipate future outcomes and future responses;
wherein the language models are also used to evaluate the desirability of projected future outcomes;
converting the language models evaluation into a format suitable to guide tree search methods;
wherein information regarding the desired goals or outcomes may be included in the current situational description;
wherein the virtual agent utilizes MCTS techniques in combination with the one or more language models for outputting textual responses to the current situation description;
wherein the virtual agent responds with textual expression to verbal input in combination with the audio, visual, tactile, and other sensory inputs after merging an analysis of basic sentiment analysis and goal oriented analysis of the AI virtual agent; and
wherein the sentiment analysis is determined by the text being tokenized into a list of words and phrases, a sentiment value is assigned for teach token in which a summation of the sentiment values are computed and then normalized as compared to an expected value to determine respective normalized scores, and then the respective normalized scores are compared to one or more indexes which are summed to provide discrete output values which provide a respective sentiment value for the text being analyzed.
11. The AI language virtual agent according to claim 9, further including the steps of:
constructing the contextual descriptions to maximally represent the current situation while meeting any processing constraints of the language models used;
providing nodes of the states of the search space which contain contextual descriptions and search control information;
receiving an initial starting contextual description and search control information; and
exploring the state of possible future contexts by using one or more language models provided with sufficient information in the form of the contextual descriptions and prompts that cause the generation of new situation descriptions.
12. The AI language virtual agent according to claim 9, further comprising:
utilizing one or more language models to generate evaluative statements or annotations on the desirability of a situational description;
translating symbolic evaluative statements into numeric or vector format suitable for use in common state space search systems;
selecting an initial action from an initial set based on the outcome of the complete search process;
performing an action or set of actions based on the initial action selected;
recording the system and environmental interactions to one or more logs;
providing for the continuing training of the language models used by the system using reference material and direct experience; and
providing self-improvement of the entire process through the iterative retraining of the language models, through the use of self-play to explore the space of novel possible situations and to discover paths towards positive resolutions in those situations.
13. The AI language virtual agent according to claim 9, wherein said virtual agent responds to verbal input in combination with various sensory inputs, including one or more of audio, visual, tactile inputs, with said combination expressed in natural language and merged with background information to create a contextual description; and
wherein in addition to textual expression the virtual language agent responds with animation and virtual effectors.
Patent Citations (21)
| Patent | Date | Inventor | Cited By |
|---|---|---|---|
| US7113848(B2) | 2006-09-01 | Hanson | Applicant |
| US7809548(B2) | 2010-10-01 | Mihalcea et al. | Applicant |
| US8346534(B2) | 2013-01-01 | Csomai et al. | Applicant |
| US8818926(B2) | 2014-08-01 | Wallace | Applicant |
| US9047423(B2) | 2015-06-01 | Tesauro et al. | Applicant |
| US9141600(B2) | 2015-09-01 | Tramp | Applicant |
| US9796095(B1) | 2017-10-01 | Hanson et al. | Applicant |
| US10008196(B2) | 2018-06-01 | Maisonnier et al. | Applicant |
| US10157342(B1) | 2018-12-01 | Kim | Applicant |
| US10272349(B2) | 2019-04-01 | Davenport | Applicant |
| US10621648(B2) | 2020-04-01 | Davis | Applicant |
| US10664741(B2) | 2020-05-01 | Gibbs et al. | Applicant |
| US20170125008(A1) | 2017-05-01 | Maisonnier et al. | Applicant |
| US20180052664(A1) | 2018-02-01 | Zhang | Examiner |
| US20190381415(A1) | 2019-12-01 | McMullen et al. | Applicant |
| US20200065868(A1) | 2020-02-01 | Garlapati et al. | Applicant |
| US20200147808(A1) | 2020-05-01 | Pirzchalski et al. | Applicant |
| US20200234243(A1) | 2020-07-01 | Miron et al. | Applicant |
| US20200258120(A1) | 2020-08-01 | Ardulov et al. | Applicant |
| US20200320978(A1) | 2020-10-01 | Chatterjee | Examiner |
| WO2018081833(A1) | 2018-05-01 | Applicant |
Non-Patent Literature (178)
- Monte-Carlo Tree Search—Chessprogramming wiki; last edit May 15, 2020; URL: https://www.chessprogramming.org/Monte-Carlo_Tree-Search (17 pages).Applicant
- Oresky, Colleen M., et al. (1991). Strategic automatic discovery system (STRADS). 4. 223-260. 10.1007/978-1-4612-3040-3_11; Reprinted from “Knowledge-Based Simulation: Methodology and Application” (38 pages).Applicant
- Orkin, J., “Goal-Oriented Action Planning (GOAP)”, MIT Media Lab, URL: http://alumni.media.mit.edu/˜jorkin/goap.html (2 pages), Nov. 9, 2020.Applicant
- Orkin, Jeff, “Applying Goal-Oriented Action Planning to Games,” (2008) URL: http://alumni.media.mit.edu/˜jorkin/GOAP_draft_AIWisdom2_2003.pdf (11 pages).Applicant
- Radford, Alec et al., “Better Language Models and Their Implications,” openai.com.pdg; (2019) URL:https://openai.com/blog/better-language-models/; (1 page).Applicant
- Radford, Alec, et al., “Language Models are Unsupervised Multitask Learners” (2019) URL: https://d4mucfpksywv.cloudfront.net/better-language-models/language_models_are_unsupervised_multitask_learners.pdf (24 pages).Applicant
- Replika: Building an Emotional Conversation with Deep Learning; URL: https://scai.info/slides/replika_ai.pdf (26 pages), Nov. 9, 2020.Applicant
- Schrittwieser, Julian, et al., “Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model,” arXiv:1911.08265v2 [cs.LG] (2020), (21 pages).Applicant
- Seth, Yashu, “BERT Explained—A list of Frequently Asked Questions” (2019) URL: https://yashuseth.blog/2019/06/12/bert-explained-faqs-understand-bert-working/ (8 pages).Applicant
- Shum, Heung-Yeung, et al., “From Eliza to Xiaolce: Challenges and Opportunities with Social Chatbots,” Frontiers of Information Technology & Electronic Engineering 19, No. 1 (2018): 10-26; URL: https://arxiv.org/ftp/arxiv/papers/1801/1801.01957.pdf (20 pages).Applicant
- Silver, David, et al., “Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm” arXiv:1712.01815 [cs.AI] (Dec. 5, 2017) URL: https://arxiv.org/pdf/1712.01815 (19 pages).Applicant
- Silver, David, et al.; “A general reinforcement learning algorithm that masters chess, shogi, and go through self-play” Science. 362 (6419); 1140-1144. doi:10.1126/science.aar6404. Retrieved Dec. 7, 2018; URL: https://science.sciencemag.org/content/sci/362/6419/1140.full.pdf (6 pages).Applicant
- Sturtevant, Nathan, “Monte-Cado Tree Search and Related Algorithms for Games”, Game AI Pro 2: Collected Wisdom of Game AI Professionals; URL: http://www.gameaipro.com/GameAIPro2/GameAIPro2_Chapter25_Monte_Carlo_Tree_Search_and_Related_Algorithms_for_Games.pdf (17 pages), Nov. 9, 2020.Applicant
- Summa 1.2.0; URL: https://pypi.org/project/summa/ (2 pages), Nov. 9, 2020.Applicant
- The Sims Video Games—Official EA Site; URL: https://www.ea.com/games/the-sims (2 pages), Nov. 9, 2020.Applicant
- The Sims, Wikipedia; URL: https://en.wikipedia.org/wiki/The_Sims (7 pages), Nov. 9, 2020.Applicant
- Transformers as Soft Reasoners over Language, AI2 Allen Institute for AI; URL: https://rule-reasoning.apps.allenai.org/ about (1 page), Nov. 9, 2020.Applicant
- UCT—Chessprogramming wiki; last edit May 15, 2020; URL: https://www.chessprogramming.org/UCT (8 pages).Applicant
- VaderSentiment 3.3.2; PyPI.org screenshot; URL: https://pypi.org/project/vaderSentiment/ (1 page), Nov. 9, 2020.Applicant
- Veness, Joel et al.; “A Monte-Cado AIXI Approximation,” Journal of Artificial Intelligence Research 40: 95-142 (2011); URL: https://www.jair.org/index.php/jair/article/view/10685 (48 pages).Applicant
- Veness, Joel, et al., “A Monte Cado AIXI Approximation,” e-print arXiv:0909.0801 (2009); URL: https://arxiv.org/abs/0909.0801 (51 pages).Applicant
- Vincent, James “DeepMind's AI became a superhuman chess player in a few hours, just for fun,” The Verge; (Dec. 6, 2017); URL: https://www.theverge.com/2017/12/6/16741106/deepmind-ai-chess-alphazero-shogi-go (4 pages).Applicant
- Wallace, Dr. Richard S. “The Elements of AIML Style,” Alice A. I. Foundation, Inc. (2003) URL: https://files.ifi.uzh.ch/cl/hess/classes/seminare/chatbots/style.pdf (86 pages), Nov. 9, 2020.Applicant
- Webb TW, Graziano MS (2015) “The attention schema theory: a mechanistic account of subjective awareness,” Front Psychol. 6:500. doi:10.3389/fpsyg.2015.00500. PMC 4407481. PMID 25954242 (11 pages).Applicant
- Xiaoice_Wikipedia, last edited Apr. 17, 2020, URL: https://en.wikipedia.org/wiki/Xiaoice (2 pages).Applicant
- Zhou, Li, et al., “The Design and Implementation of Xiaoice, an Empathetic Social Chatbot,” Computational Linguistics 46, No. 1 (2020); 53-93; URL: https://arxiv.org/pdf/1812.08989.pdf (35 pages).Applicant
- Abstract Meaning Representation (AMR)-Home; URL: https://amr.isi.edu/index.html (1 page), Nov. 9, 2020.Applicant
- Abstract Meaning Representation (AMR)-Research; URL: https://amr.isi.edu/research.html (3 pages), Nov. 9, 2020.Applicant
- Alammar, Jay, “The Illustrated GPT-2 (Visualizing Transformer Language Models)” Visualizing machine learning one concept at a time; URL: http//jalammar.github.io/illustrated-gpt2/ (13 pages), Nov. 9, 2020.Applicant
- Alammar, Jay, “The Illustrated BERT, ELMo, and co. (How NLP Cracked Transfer Learning)” (2018); URL: http://alammar.github.io/illustrated-bert/ (20 pages).Applicant
- BOINC Compute for Science, University of California; URL: https://boinc.berkeley.edu/ (1 page), Nov. 9, 2020.Applicant
- Bourse, Yoann, “Artificial Intelligence in the Sims Series”; URL: https://team.inria.fr/imagine/files/2014/10/sims-slides.pdf (44 pages).Applicant
- Brown, Tom B., et al.; “Language models are few-shot learners,” arXiv preprint arXiv:2005.14165 (2020); URL: https://arxiv.org/pdf/2005.14165 and URL: https://github.com/openai/gpt-3 (74 pages).Applicant
- Browne, Cameron B., et al., “A Survey of Monte Carlo Tree Search Methods”, IEEE Transactions on Computational Intelligence and AI in Games, vol. 4, No. 1, Mar. 2012 (Mar. 1, 2012) pp. 1-43 (49 pages).Applicant
- Chan, David, “The philosophy of The Sims,” Stanford University Working paper; http://library.stanford.edu/depts/hasrg/histsci/STS145papers/Chan. pdf (2003);URL: https://web.stanford.edu/group/htgg/cgi-bin/drupal/sites/default/files2/dchan_2003_1.pdf (12 pages).Applicant
- Clark, Peter, et al., “Transformers as Soft Reasoners over Language.” ArXiv abs/2002.05867 (2020) n. pag. (15 pages).Applicant
- CloudTurbine Stream Sharing; URL: https://www.cloudturbine.com/ (3 pages), Nov. 9, 2020.Applicant
- Coursey, K. (2020) Speaking with Harmony: Finding the right thing to do or say . . . while in bed (or anywhere else). In O. Bendel (Ed.), Maschinenliebe: Teil 3 Liebespuppen und Sexroboter aus technischer, psychologischer und philosophischer Perspektive (pp. 1-12). Wiesbaden: Springer Gabler (12 pages).Applicant
- Coursey, K. (2020) Speaking with Harmony: Finding the right thing to do or say . . . while in bed (or anywhere else). In O. Bendel (Ed.), Maschinenliebe: Teil 1 Liebespuppen und Sexroboter aus technischer, psychologischer und philosophischer Perspektive. Wiesbaden: Springer Gabler (17 pages).Applicant
- Coursey, K., et al., (2019) “Living with Harmony: A Personal Companion System by Realbotix™.” In: Zhou Y., Fischer MH (eds) AI Love You. Springer, Cham, pp. 77-95; URL: https://www.daxtron.com/pdf/Coursey%20et%20aI.%20Living%20with%20Harmony%20AI+Love+You.pdf (36 pages).Applicant
- Coursey, Kino High, “The Value of Everything: Ranking and Association with Encyclopedic Knowledge,” University of North Texas (2009); URL: https://digital.library.unt.edU/ark:/67531/metadc12108/m2/1/high_res_d/dissertaLion.pdf (191 pages).Applicant
- Csomai, A. (2008) “Keywords in the mist: Automated keyword extraction for very large documents and back of the book indexing,” Ph.D. thesis, University of North Texas, United States—Texas; URL: https://digital.library.unt.edu/ark:/67531/metadc6118/m2/1/high_res_d/dissertation.pdf (145 pages).Applicant
- Csomai, A. and Mihalcea, R. (2008) “Linking Documents to Encyclopedic Knowledge,” IEEE Intelligent Systems 23 (5), 34-41; URL: http://www.academia.edu/download/8026958/mex2008050034.pdf (8 pages).Applicant
- Devlin, J. et al., “Bert: Pre-training of deep bidirectional transformers for language understanding” arXiv preprint arXiv . . . , (2018) arxiv.org; URL: https://arxiv.org/pdf/1810.04805.pdf (16 pages).Applicant
- FightAIDS@home, “Go to battle against AIDS with your computer!” URL: http://fightaidsathome.scripps.edu/ (3 pages), Nov. 9, 2020.Applicant
- Folding@Home; URL: https://foldingathome.org/2019/03/?lng=en-US (1 page).Applicant
- Gao, Jianfeng, et al., “Neural Approaches to Conversational AI”, SIGIR (2018); URL: https://www.microsoft.com/en-us/research/uploads/prod/2018/07/neural-approaches-to-conversational-AI.pdf (175 pages).Applicant
- GPT-2 Neural Network Poetry; gwern.net; URL: https://www.gwem.net/GPT-2 (88 pages), Nov. 9, 2020.Applicant
- Grand, Stephen and Dave CLIFF, “Creatures: Entertainment Software Agents with Artificial Life,” Autonomous Agents and Multi-Agent Systems, 100, 1-20 (1997); URL: https://www.researchgate.net/publication/226997131 (22 pages).Applicant
- Grand, Steve, “Creation: Life and How to Make It” (2000) (173 pages).Applicant
- Grand, Steve, “Growing Up With Lucy: How to Build an Android in Twenty Easy Steps,” (2003) (214 pages).Applicant
- Grand, Steve, et al., “Creatures: Artificial Life Autonomous Software Agents for Home Entertainment,” Millennium Technical Report 9601, University of Sussex Technical Report CSRP434 (1997) (9 pages).Applicant
- Grand, Steven and Dave Cliff, “Creatures: Entertainment Software Agents with Artificial Life,” Autonomous Agents and Multi-Agent Systems, 100, 1-20 (1997) URL: https://www.researchgate.net/publication/220794226_Creatures_Artificial_Life_Autonomous_Software_Agents_for_Home_Entertainment; Fall 2005 (14 pages).Applicant
- Graziano, M.S.A., Webb, T.W. (2017) “From Sponge to Human: The Evolution of Consciousness” (PDF). Evolution of Nervous Systems, 2E. vol. 3, pp. 547-554. doi:10.1016/B978-0-12-804042-3.00098-1. ISBN 9780128040966; URL: http://www.elsevier.com/locate/permissionusematerial (9 pages), Nov. 9, 2020.Applicant
- Graziano, Michael and Taylor Webb, “Part Three, Metaphilosophy of Consciousness Studies, Understanding Consciousness by Building It,” The Bloomsbury Companion to the Philosophy of Consciousness, (2018) 185-210 (26 pages).Applicant
- Graziano, MSA (2017) “The Attention Scheme Theory: A Foundation for Engineering Artificial Consciousness,” Frontiers in Robotics and AI 4:60. doi: 10.3389/frobt.2017.00060 (2017) (9 pages).Applicant
- Haase, K. W. (1990) “Invention and Exploration in Discovery,” (Doctoral dissertation, Massachusetts Institute of Technology);URL: https://dspace.mit.edu/bitstream/handle/1721.1/14257/22713693-MIT.pdf?sequence=2 (165 pages), Nov. 9, 2020.Applicant
- Hopkins, Don, “The Sims Design Document” v.71 “The Sims Design Document Draft 7” (Oct. 2, 1998) URL: https://donhopkins.com/home/TheSims/TheSimsDesignDocumentDraft7-1 998-10-02-DonsReview.pdf; Assoc URL: http//linkedbyair.net/bin/The%20Sims%20Design%20Documents/(51 pages).Applicant
- Horev, Rani, “BERT Explained: State of the art language model for NLP” (2018); URL: https://towardsdatascience.com/bert-explained-state-of-the-art-language-model-for-nlp-f8b21a9b6270 (8 pages).Applicant
- How to Train It, Issue #19, GitHub, Feb. 15, 2019; URL: https://github.com/openai/gpt-2/issues/19 (6 pages).Applicant
- Hutto, Clayton J. and Eric Gilbert, “VADER: A Parsimonious Rule-Based Model for Sentiment Analysis of Social Media Text ” Eighth International Conference on Weblogs and Social Media (ICWSM-14) Jun. 2014 (10 pages).Applicant
- Install TensorFlow 2; URL: https://www.tensorflow.org/install (1 page), Nov. 9, 2020.Applicant
- Install TensorFlow with pip; URL: https://www.tensorflow.org/install/pip (3 pages), Nov. 9, 2020.Applicant
- Kocsis, Levente, et al., “Bandit based Monte-Carlo Planning” (2006); In Furnkranz, Johannes; Scheffer, Tobias; Spiliopoulou, Myra (eds ). Machine Learning: ECML 2006, 17th European Conference on Machine Learning, Berlin, Germany (Sep. 18-22, 2006), Proceedings. Lecture Notes in Computer Science. 4212. Springer, pp. 282-293. CiteSeerX 10.1.1.102.1296. doi:10.1007/11871842_29. ISBN 3-540-45375-X; Url: https://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.102.1296 (12 pages).Applicant
- Lee, Jieh-Sheng and Hsiang, Jieh, “Patent Claim Generation by Fine-Tuning OpenAI GPT-2.” arXiv preprint arXiv: 1907.02052 (2019); National Taiwan University; URL: https://arxiv.org/pdf/1907.02052 (11 pages).Applicant
- Lenat, Douglas (1983) “EURISKO: A program that learns new heuristics and domain concepts,” Artificial Intelligence 21 (1-2): 61-98, doi:10.1016/s0004-3702(83)80005-8 (39 pages), Nov. 9, 2020.Applicant
- Lenat, Douglas and Brown, J.S. (1984) “Why AM and EURISKO appear to work” (PDF), ResearchGate: 269-240; URL: https://www.researchgate.net/publication/221605660_Why_AM_and_Eurisko_Appear_to_Work (6 pages), Nov. 9, 2020.Applicant
- Lenat, Douglas B., “The Nature of Heuristics,” Artiicial Intelligence 19 (1982) 189-249 (61 pages).Applicant
- Lenat, Douglas B., “Theory Formation by Heuristic Search,” Artificial Intelligence 21 (1983) 31-59 (29 pages).Applicant
- Lenat, Douglas, “EURISKO: A program that learns new heuristics and domain concepts,” Artificial Intelligence 21 (1983) 61-98, doi:10 1016/50004-3702(83)80005-8 (38 pages).Applicant
- Lenat, Douglas; Brown, JS (1984) “Why AM and EURISKO Appear to Work,” Artificial Intelligence 23(3), 269-294. https://doi.org/10.1016/0004-3702(84)90016-X (26 pages), Nov. 9, 2020.Applicant
- Mcts 1.0.4; URL: https://pypi.org/project/mcts/ (1 page), Nov. 9, 2020.Applicant
- Microsoft_AI_Vector_search; URL: https://www.microsoft.com/en-us/ai/ai-lab-vector-search (4 pages), Nov. 9, 2020.Applicant
- Mihalcea, Rada and Csomai, Andras, “Wikifyl Linking documents to encyclopedic knowledge,” In Proceedings of the sixteenth ACM conference on Conference on information and knowledge management, pp. 233-242. 2007; URL: http://web.eecs.umich.edu/˜mihalcea/papers/mihalcea.cikm07.pdf (9 pages).Applicant
- Mihalcea, Rada and Tarau, Paul (2004) TextRank: “Bringing Order into Texts,” Department of Computer Science University of North Texas “Archived copy” (PDF). Archived from the original (PDF) on Jun. 17, 2012. Retrieved Jul. 20, 2012; URL: https://web.archive.org/web/20120617170501/http:/acl.ldc.upenn.edu/acl2004/emnlp/pdf/Mihalcea.pdf (8 pages).Applicant
- Monte Carlo tree search—Wikipedia; last edit Jun. 10, 2020; URL: https://en.wikipedia.org/wiki/Monte_Carlo_tree_search (10 pages).Applicant
- 48_Source_Code_GitClone_gpt-2_801-1000, Nov. 9, 2020.Applicant
- 49_Source_Code_GitClone_gpt-2_1001-1200, Nov. 9, 2020.Applicant
- 50_Source_Code_GitClone_gpt-2_1201-1305, Nov. 9, 2020.Applicant
- 51_Source_Code_GitClone_gpt-2-simple_1-200, Nov. 9, 2020.Applicant
- 52_Source_Code_GitClone_gpt-2-simple_201-400, Nov. 9, 2020.Applicant
- 53_Source_Code_GitClone_gpt-2-simple_401-600, Nov. 9, 2020.Applicant
- 54_Source_Code_GitClone_gpt-2-simple_601-800, Nov. 9, 2020.Applicant
- 55_Source_Code_GitClone_gpt-2-simple_801-1000, Nov. 9, 2020.Applicant
- 56_Source_Code_GitClone_gpt-2-simple_1001-1200, Nov. 9, 2020.Applicant
- 57_Source_Code_GitClone_gpt-2-simple_1201-1440, Nov. 9, 2020.Applicant
- 58_Source_Code_GitClone_gpt-3_1-500, Nov. 9, 2020.Applicant
- 59_Source_Code_G itClone_gpt-3_501-1000, Nov. 9, 2020.Applicant
- 60_Source_Code_GitClone_gpt-3_1001-1500, Nov. 9, 2020.Applicant
- 61_Source_Code_GitClone_gpt-3_1501-2000, Nov. 9, 2020.Applicant
- 62_Source_Code_GitClone_gpt-3_2001-2500, Nov. 9, 2020.Applicant
- 63_Source_Code_GitClone_gpt-3_2501-3539, Nov. 9, 2020.Applicant
- 64_Source_Code_G itClone_johan-gras_136, Nov. 9, 2020.Applicant
- 65_Source_Code_GitClone_MCTS_1-50, Nov. 9, 2020.Applicant
- 66_Source_Code_GitClone_MCTS_51-98, Nov. 9, 2020.Applicant
- 67_Source_Code_GitClone_milvus_1-700, Nov. 9, 2020.Applicant
- 68_Source_Code_GitClone_milvus_701-2100, Nov. 9, 2020.Applicant
- 69_Source_Code_GitClone_milvus_2101-3500, Nov. 9, 2020.Applicant
- 70_Source_Code_GitClone_milvus_3501-4200, Nov. 9, 2020.Applicant
- 71_Source_Code_GitClone_milvus_4201-4900, Nov. 9, 2020.Applicant
- 72_Source_Code_GitClone_milvus_4901-5600, Nov. 9, 2020.Applicant
- 73_Source_Code_GitClone_milvus_5601-6300, Nov. 9, 2020.Applicant
- 74_Source_Code_GitClone_milvus_6301-6848, Nov. 9, 2020.Applicant
- 75_Source_Code_GitClone_muzero-general_1-1600, Nov. 9, 2020.Applicant
- 76_Source_Code_GitClone_muzero-general_1601-1983, Nov. 9, 2020.Applicant
- 77_Source_Code_GitClone_muzero-pytorch_556, Nov. 9, 2020.Applicant
- 78_Source_Code_GitClone_pynndescent_1-400, Nov. 9, 2020.Applicant
- 79_Source_Code_GitClone_pynndescent_401-800, Nov. 9, 2020.Applicant
- 80_Source_Code_GitClone_python-tf-idf_114, Nov. 9, 2020.Applicant
- 81_Source_Code_GitClone_ReDial_A19_1-1000, Nov. 9, 2020.Applicant
- 82_Source_Code_GitClone_ReDial_A19_1001-3000, Nov. 9, 2020.Applicant
- 83_Source_Code_GitClone_ReDial_A19_3001-5000, Nov. 9, 2020.Applicant
- 84_Source_Code_GitClone_ReDial_A19_5001-7000, Nov. 9, 2020.Applicant
- 85_Source_Code_GitClone_ReDial_A19_7001-7585, Nov. 9, 2020.Applicant
- 86_Source_Code_GitClone_ReducedFreeSims_1-600, Nov. 9, 2020.Applicant
- 87_Source_Code_GitClone_ReducedFreeSims_601-1200, Nov. 9, 2020.Applicant
- 88_Source_Code_GitClone_ReducedFreeSims_1201-4800, Nov. 9, 2020.Applicant
- 89_Source_Code_GitClone_ReducedFreeSims_4801-8400, Nov. 9, 2020.Applicant
- 90_Source_Code_GitClone_ReducedFreeSims_8401-10800, Nov. 9, 2020.Applicant
- 91_Source_Code_GitClone_ReducedFreeSims_10801-12600, Nov. 9, 2020.Applicant
- 92_Source_Code_GitClone_ReducedFreeSims_12601-13800, Nov. 9, 2020.Applicant
- 93_Source_Code_GitClone_ReducedFreeSims_13801-15600, Nov. 9, 2020.Applicant
- 94_Source_Code_GitClone_ReducedFreeSims_15601-16800, Nov. 9, 2020.Applicant
- 95_Source_Code_GitClone_ReducedFreeSims_16801-17717, Nov. 9, 2020.Applicant
- 96_Source_Code_GitClone_SPTAG_561, Nov. 9, 2020.Applicant
- 97_Source_Code_GitClone_textrank_251, Nov. 9, 2020.Applicant
- 98_Source_Code_GitClone_trainLMS_pub_9, Nov. 9, 2020.Applicant
- 99_Source_Code_GitClone_vaderSentiment_1-750, Nov. 9, 2020.Applicant
- 01_Source_Code_GitClone_AI-DungeonMaster_1-100, Nov. 9, 2020.Applicant
- 02_Source_Code_GitClone_AI-DungeonMaster_101-781, Nov. 9, 2020.Applicant
- 03_Source_Code_GitClone_ann-benchmarks_292, Nov. 9, 2020.Applicant
- 04_Source_Code_GitClone_annoy_1-600, Nov. 9, 2020.Applicant
- 05_Source_Code_GitClone_annoy_601-1105, Nov. 9, 2020.Applicant
- 06_Source_Code_GitClone_cdQA_2384, Nov. 9, 2020.Applicant
- 07_Source_Code_GitClone_cloudturbine_1-3000, Nov. 9, 2020.Applicant
- 08_Source_Code_GitClone_cloudturbine_3001-6000, Nov. 9, 2020.Applicant
- 09_Source_Code_GitClone_cloudturbine_6001-9000, Nov. 9, 2020.Applicant
- 10_Source_Code_GitClone_cloudturbine_9001-12000, Nov. 9, 2020.Applicant
- 100_Source_Code_GitClone_vaderSentiment_751-1412, Nov. 9, 2020.Applicant
- 101_Source_Code_GitClone_Zeta36_1-10, Nov. 9, 2020.Applicant
- 102_Source_Code_GitClone_Zeta36_11-131, Nov. 9, 2020.Applicant
- 11_Source_Code_GitClone_cloudturbine_12001-15000, Nov. 9, 2020.Applicant
- 12_Source_Code_GitClone_cloudturbine_15001-18000, Nov. 9, 2020.Applicant
- 13_Source_Code_GitClone_cloudturbine_18001-21000, Nov. 9, 2020.Applicant
- 14_Source_Code_GitClone_cloudturbine_21001-24000, Nov. 9, 2020.Applicant
- 15_Source_Code_GitClone_cloudturbine_24001-27000, Nov. 9, 2020.Applicant
- 16_Source_Code_GitClone_cloudturbine_27001-30000, Nov. 9, 2020.Applicant
- 17_Source_Code_GitClone_cloudturbine_30001-33000, Nov. 9, 2020.Applicant
- 18_Source_Code_GitClone_cloudturbine_33001-36000, Nov. 9, 2020.Applicant
- 19_Source_Code_GitClone_cloudturbine_36001-39000, Nov. 9, 2020.Applicant
- 20_Source_Code_GitClone_cloudturbine_39001-42000, Nov. 9, 2020.Applicant
- 21_Source_Code_GitClone_cloudturbine_42001-45000, Nov. 9, 2020.Applicant
- 22_Source_Code_GitClone_cloudturbine_45001-48000, Nov. 9, 2020.Applicant
- 23_Source_Code_GitClone_cloudturbine_48001-51000, Nov. 9, 2020.Applicant
- 24_Source_Code_GitClone_cloudturbine_51001-54000, Nov. 9, 2020.Applicant
- 25_Source_Code_GitClone_cloudturbine_54001-57000, Nov. 9, 2020.Applicant
- 26_Source_Code_GitClone_cloudturbine_57001-60000, Nov. 9, 2020.Applicant
- 27_Source_Code_GitClone_cloudturbine_60001-63000, Nov. 9, 2020.Applicant
- 28_Source_Code_GitClone_cloudturbine_63001-66000, Nov. 9, 2020.Applicant
- 29_Source_Code_GitClone_cloudturbine_66001-66882, Nov. 9, 2020.Applicant
- 30_Source_Code_GitClone_fast-bert_1-3000, Nov. 9, 2020.Applicant
- 31_Source_Code_GitClone_fast-bert_3001-4587, Nov. 9, 2020.Applicant
- 32_Source_Code_GitClone_Fine-tune_7, Nov. 9, 2020.Applicant
- 33_Source_Code_GitClone_FreeSims_1-700, Nov. 9, 2020.Applicant
- 34_Source_Code_GitClone_FreeSims_701-2800, Nov. 9, 2020.Applicant
- 35_Source_Code_GitClone_FreeSims_2801_6300, Nov. 9, 2020.Applicant
- 36_Source_Code_GitClone_FreeSims_6301-9100, Nov. 9, 2020.Applicant
- 37_Source_Code_GitClone_FreeSims_9101-11900, Nov. 9, 2020.Applicant
- 38_Source_Code_GitClone_FreeSims_11901-14000, Nov. 9, 2020.Applicant
- 39_Source_Code_GitClone_FreeSims_14001-15400, Nov. 9, 2020.Applicant
- 40_Source_Code_GitClone_FreeSims_15401-17500, Nov. 9, 2020.Applicant
- 41_Source_Code_GitClone_FreeSims_17501-18417, Nov. 9, 2020.Applicant
- 42_Source_Code_GitClone_GitClone_AI-DungeonMaster_1-300, Nov. 9, 2020.Applicant
- 43_Source_Code_GitClone_GitClone_AI-DungeonMaster_301-781, Nov. 9, 2020.Applicant
- 44_Source_Code_GitClone_gpt-2_1-200, Nov. 9, 2020.Applicant
- 45_Source_Code_GitClone_gpt-2_201-400, Nov. 9, 2020.Applicant
- 46_Source_Code_GitClone_gpt-2_401-600, Nov. 9, 2020.Applicant
- 47_Source_Code_GitClone_gpt-2_601-800, Nov. 9, 2020.Applicant