Adapting Bayesian network parameters on-line in a dynamic environment

[0001] 1. Field of Invention

[0002] The present invention pertains to the field of automated reasoning. More particularly, this invention relates to Bayesian networks in automated reasoning.

[0003] 2. Art Background

[0004] Bayesian networks are commonly used for automated reasoning in a wide variety of applications. Typically, Bayesian networks are used to model an underlying system or environment of interest. For example, Bayesian networks may be used to model biological systems including humans and animals, electrical systems, mechanical systems, software systems, business transaction systems, etc. Bayesian networks may be useful in a variety of automated reasoning tasks including diagnosing problems with an underlying system, determining the health of an underlying system, and predicting future events in an underlying system to name a few examples.

[0005] A typical Bayesian network is a graph structure having a set of nodes and interconnecting arcs that define parent-child relationships among the nodes. A Bayesian network also includes a set of Bayesian network parameters which are associated with the nodes of the graph structure. Typically, the nodes of a Bayesian network are associated with events or characteristics of the underlying modeled environment and the Bayesian network parameters usually indicate partial causalities among the events or characteristics associated with the nodes. The Bayesian network parameters are commonly contained in conditional probability tables associated with the nodes. Typically, a Bayesian network describes the joint probability of random variables each of which is represented by a node in the graph.

[0006] The Bayesian network parameters are commonly obtained from experts who possess knowledge of the behaviors or characteristics of the underlying modeled environment. Alternatively, the Bayesian network parameters may be obtained using observations of the underlying modeled environment. Unfortunately, environments may exist for which experts are unavailable or prohibitively expensive. In addition, environments may exist for which observation data is scarce or in which the underlying environment changes and renders past experience or observations obsolete.

[0007] A method is disclosed for adapting a Bayesian network. A Bayesian network may be adapted to changes in the underlying modeled environment using the present techniques even when observation data is relatively scarce and in an on-line dynamic environment. The present method includes determining a set of parameters for the Bayesian network, for example, initial parameters, and then updating the parameters in response to a set of observation data using an adaptive learning rate. The adaptive learning rate responds to any changes in the underlying modeled environment using minimal observation data.

[0008] Other features and advantages of the present invention will be apparent from the detailed description that follows.

[0009] The present invention is described with respect to particular exemplary embodiments thereof and reference is accordingly made to the drawings in which:

[0010] FIG. 1 shows an on-line adapter which adapts a Bayesian network according to the present teachings;

[0011] FIG. 2 shows a method for adapting Bayesian network parameters according to the present teachings;

[0012] FIG. 3 shows an example Bayesian network for illustrating the present techniques;

[0013] FIG. 4 shows a method for determining an adaptive learning rate according to the present teachings.

[0014] FIG. 1 shows an on-line adapter 56 which adapts a Bayesian network 52 according to the present teachings. The Bayesian network 52 performs automated reasoning with respect to an on-line environment 50. The on-line adapter 56 obtains a set of observation data 54 from one or more elements of the on-line environment 50 and adapts the parameters of the Bayesian network 52 in response to the observation data 54.

[0015] The present techniques enables adaptation of the parameters of the Bayesian network 52 in response to changes in the on-line environment 50 even when the observation data 54 is relatively scarce and/or when some of values in the observation data 54 from elements of the on-line environment 50 are unavailable. For example, at any given time values from a subset of the hardware/software elements of the on-line environment 50 may be unavailable due to hardware/software failures and/or due to the nature of events in the on-line environment 50.

[0016] The on-line environment 50 may be the hardware/software elements of an email system, an e-commerce system, a database system, or any type of distributed application to name a few examples.

[0017] FIG. 2 shows a method for adapting Bayesian network parameters according to the present teachings. At step 102, a set of initial Bayesian network parameters are determined. At step 104, the Bayesian network parameters are updated in response to a set of observation data using an adaptive learning rate. The step 104 may be repeated for each of a set of observation data records.

[0018] FIG. 3 shows an example Bayesian network 100 for illustrating the present techniques. The Bayesian network 100 includes a set of nodes 10-14. The node 10 corresponds to a variable (EXCHANGE_A) which indicates whether a stock exchange A is up (U) or down (D) in terms of change in value. The node 12 corresponds to a variable (EXCHANGE_B) which indicates whether a stock exchange B is up or down. The node 14 corresponds to a variable (MY_STOCKS) which indicates whether the value of a set of stocks associated with the stock exchanges A and B and held by a particular individual is up or down.

[0019] The nodes 10-14 have associated conditional probability tables 20-24, respectively, for holding the parameters of the Bayesian network 100. The conditional probability tables 20-24 are written with a set of initial Bayesian network parameters determined at step 102.

[0020] An example set of initial Bayesian network parameters in the conditional probability table 20 is as follows:

entry	D	U
0	0.2	0.8

[0021] Entry 0 in the conditional probability table 20 indicates that the probability that the variable EXCHANGE_AD is 20 percent and the probability that the variable EXCHANGE_AU is 80 percent under all conditions.

[0022] An example set of initial Bayesian network parameters in the conditional probability table 22 is as follows:

entry	D	U
0	0.3	0.7

[0023] Entry 0 in the in the conditional probability table 22 indicates that the probability that the variable EXCHANGE_BD is 30 percent and the probability that the variable EXCHANGE_BU is 70 percent under all conditions.

[0024] An example set of initial Bayesian network parameters in the conditional probability table 24 are as follows:

entry	EXCHANGE_A	EXCHANGE_B	D	U
0	D	D	0.5	0.5
1	D	U	0.5	0.5
2	U	D	0.5	0.5
3	U	U	0.5	0.5

[0025] Entry 0 in the in the conditional probability table 24 indicate that the probability that the variable MY_STOCKSD is 50 percent and that the probability that the variable MY_STOCKSU is 50 percent given that EXCHANGE_AD and EXCHANGE_BD. Similarly, entries 1-3 in the conditional probability table 24 indicate that the probability that the variable MY_STOCKSD is 50 percent and that the probability that the variable MY_STOCKSU is 50 percent given that all remaining combinations of conditions of EXCHANGE_A and EXCHANGE_B.

[0026] At step 104, the Bayesian network parameters contained in the conditional probability table 24 are updated in response to the following example set of observation data:

[0027] EXCHANGE_AU

[0028] EXCHANGE_BD

[0029] MY_STOCKSD

[0030] This example observation data corresponds to entry 2 in the conditional probability table 24 with the conditions EXCHANGE_AU and EXCHANGE_BD.

[0031] Entry 2 is updated at step 104 in response to the example observation as follows. The new probability that MY_STOCKSD given that EXCHANGE_A U and EXCHANGE_BD equals (1q) times the previous probability that MY_STOCKSD given that EXCHANGE_AU and EXCHANGE_BD, where is an adaptive learning rate which is a number between 0 and 1. This increases the probablity that MY_STOCKS D given that EXCHANGE_AU AND EXCHANGE_BD.

[0032] In addition, the probablity that MY_STOCKSU given that EXCHANGE_AU and EXCHANGE_BD is decreased at step 104 as follows. The new probability that MY_STOCKSU given that EXCHANGE_A U and EXCHANGE_BD equals (1) times the previous probability that MY_STOCKSU given that EXCHANGE_AU AND EXCHANGE_BD.

[0033] The updated conditional probability table 24 after step 104 is as follows:

entry	EXCHANGE_A	EXCHANGE_B	D	U
0	D	D	0.5	0.5
1	D	U	0.5	0.5
2	U	D	(1 )	(1 ) 0.5
			0.5
3	U	U	0.5	0.5

[0034] The following generalized description enables adaptation for systems including on-line environments in which some of the values in the observation data 54 are unavailable. In the following, Z_i is a node in a Bayesian network that takes any value from the set:

[0035] Pa₁ is the set of parents of Z_i in the Bayesian network that take on of the configurations denoted by the following:

[0036] For the Bayesian network 100 example, if Z₁ is the node 14 then Pa₁ and Pa₂ are the nodes 10 and 12 and the configurations pa_i¹, . . . are DD, DU, UD, and

[0037] An entry in the conditional probability table for the node Z_i is given by the following:

[0038] A set of observation data cases D are represented as follows:

[0039] The update of the parameters for the conditional probability table for the node Z_i is achieved by the following maximization (equation 1):

[0040] where L_D () is the normalized log likelihood of the data given the network, d(,{overscore ()}) is a distance between the two models and is the learning rate. In one embodiment, the distance is the Chi squared distance. The maximization is solved under the constraint that _kijk1 for i,j.

[0041] For each new observation vector, the parameters for all conditional probability tables in a Bayesian network may be updated according to the following (equation 2): ${}_{\mathrm{ijk}}^{}={}_{\mathrm{ijk}}^{-1}+\left[\frac{P(z_i^k,{\mathrm{pa}}_i^{j}y,{}^{-1})}{P({\mathrm{pa}}_i^{j}y,{}^{-1})}-{}_{\mathrm{ijk}}^{-1}\right]$

[0042] for

[0043] and

[0044] otherwise.

[0045] This update process may be referred to as stochastic learning in which the term $\frac{P(z_i^k,{\mathrm{pa}}_i^{j}y,{}^{-1})}{P({\mathrm{pa}}_i^{j}y,{}^{-1})}$

[0046] is an instantaneous gradient estimate of the constraint optimization problem.

[0047] The learning rate may be used to control the amount of reliance on past observation data when updating the conditional probability table parameters of a Bayesian network. As approaches 1 the past observation data is weighted less and the update of the parameters is based more on the present observation data. As approaches 0 the parameters are updated slowly from the previous parameters.

[0048] The update rule of equation 2 may be rewritten as follows (equation 3) assuming a constant learning rate :

[0049] assuming with no loss of generality that

[0050] for all t1, . . . T, . . ., i.e., the parents are always observed in their j_th configuration. The assumption is a notational convention and not an actual restriction or constraint.

[0051] I_t is an indicator function and the process I_t an independent identically distributed Bernoulli random process given as I_t1 with probability _ijkc* and I_t0 with probability 1c* (equation 4) where

[0052] is the true conditional probability table entry of the Bayesian network.

[0053] Given a discrete Bayesian network S, a sequence of full observation vectors D, the update rule given in equation 3, and the constraint 0<<1, it may be shown that the following properties hold:

[0054] Property 1; X_t is a consistent estimate of c*, i.e.,

[0055] where X₀ is the initial value set at t0.

[0056] Property 2; The variance of the estimate X_t is finite and follows (equations 6 and 7): $\mathrm{Var}\left[X_t\right]=\frac{c^{*}\left(1-c^{*}\right)}{2-}\left(1-{\left(1-\right)}^{2t+2}\right){\lim }_t\mathrm{Var}\left[X_t\right]=\frac{}{2-}{c}^{*}\left(1-c^{*}\right)$

[0057] Property 3; For tax the following inequality holds: $P\left(X_t-c^{*}q\right)\frac{1}{q^2}$

[0058] for any q>0.

[0059] It is apparent that in the mean the online update rule of equation 3 approaches true values for Bayesian network parameters. The learning rate controls the rate of convergence. A setting of 1 yields the fastest convergence with the largest variance. Smaller values of yields slower convergence with smaller variance. The variance is proportional to and remains finite in the limit and thus the estimate for a Bayesian network parameter oscillates around its true value. The learning rate may be viewed as a forgetting bias of the learning algorithm so that the system forgets the past at an exponential rate proportional to .

[0060] Property 3 provides the confidence intervals of an estimated Bayesian network parameter with respect to the variance of the estimate. Property 3 may be employed in adapting the learning rate to changes in the underlying modeled environment.

[0061] FIG. 4 shows a method for adapting the learning rate according to the present teachings. At step 112, the learning rate is initially set to a value between 0 and 1. The learning rate in one embodiment is set to a relatively high value at step 112.

[0062] At step 114, an estimate of the Bayesian network parameters for an underlying modeled environment is determined in response to a set of observation data. The estimate of the Bayesian network parameters may be obtained using equation 2 given above.

[0063] At step 116, the learning rate is increased if an error between the estimate obtained at step 114 and a mean value of the Bayesian network parameters for the underlying modeled environment is relatively large. A relatively large error may occur, for example, when the modeled underlying environment changes. In one embodiment, a large error is indicated by the inequality stated in property 3 above.

[0064] Otherwise at step 118, the learning rate is decreased when convergence is reached between the estimate determined at step 114 and the mean value of the Bayesian network parameters for the underlying modeled environment.

[0065] In the following example embodiment, a different learning rate is assigned for each set of parameters of a Bayesian network, i.e. for each node Z_i with parents Pa_i and parent's configuration pa_i^j the learning rate is denoted as _ij. Letting T denote the total number of observation data, t denote the number of times Pa_ipa_i^j, and t denote the number of times Pa₁pa_i^j since the last time _ij, the schedule for adapting the learning rate is as follows.

[0066] Initialize the following:

[0067] Set PX_ix_i^kPa_ipa_i^jt_ijk to an initial value for k1, . . . ,r_i.

[0068] Set _ij to a value between 0 and 1. For example a high value may be initially set.

[0069] Set t, t0.

[0070] Given an observation vector Y_T, if Pa₁pa₁^j then do the following:

[0071] Step 1: Estimate ^t1_ijk using equation 3 where is replaced by _ij.

[0072] Step 2: If ^t1_ijkE^t1_ijk>q^t1_ij then increase _ij: set _ijijm, m>1 set t0

[0073] Else if (1_1j)^t<<1 decrease _ij: set _ijijm¹ set t0

[0074] Else set tt1

[0075] Step 3: Get the next observation and repeat steps 1-2.

[0076] The parameters q, m, and are adjustable. The parameter q determines the confidence in the decision to increase . The parameter is a threshold reflecting the acceptable convergence of the parameters. E^t1_ijk and ^t1_ij are the mean and variance of the estimated parameter. The mean may be estimated by a running average to be reset each time _ij is increased. The variance may be estimated according to equation 6 with t replaced by t and with c*0.5 which provides the worst case estimate of the variance.

[0077] It may be shown that the rate of decrease of _ij is proportional to 1/t_n where t_n is the number of times Pa_ipa₁^j until the n'th reduction of . This rate is consistent with an optimal annealing rate 1/t. In the following analysis it is assumed that _ij is not increased at any point in time. If _ij is increased, t is reset to zero and the analysis still holds. Using the reduction rule outlined above, with _ij(0)₀ and for t_ntt_n1 the learning rate _ij(t) is bounded by the following: $\frac{\log \left({}^{-1}\right)}{m-1}\frac{1}{t_n+K+\log \left({}^{-1}\right)n}{}_{\mathrm{ij}}\left(t\right)<\frac{\log \left({}^{-1}\right)}{m-1}\frac{1}{t_n-n+K}$ $\mathrm{where}$ $K=\frac{\log \left({}^{-1}\right)}{n_0\left(m-1\right)},0<<1,m>1,n.$

[0078] The bounds become tighter as t_n increases. _ij is reduced at discrete steps which increase in length as t increases. Therefore, _ij will have longer intervals at which it remains constant but at the end of the interval it reduces as 1/t. When the error between the current estimate and its mean value remains small, _ij reduces with the optimal schedule. At the limit the estimated Bayesian network parameters converge to the target parameters with zero-error. If the error becomes large, _ij increases which increases the ability to adapt faster to changes in the modeled environment or break out of a local maxima. The time origin is effectively shifted forward each time _ij is increased, which yields a method that is substantially insensitive to an absolute time origin.

[0079] The foregoing detailed description of the present invention is provided for the purposes of illustration and is not intended to be exhaustive or to limit the invention to the precise embodiment disclosed. Accordingly, the scope of the present invention is defined by the appended claims.