Method for handoff in wireless communication systems using pattern recognition

[0002] The present invention relates to wireless communication systems. More specifically, it relates to improved handoff techniques in cellular communication networks.

[0003] Wireless communication networks provide communication services to mobile devices such as cellular telephones, wireless modems, and pagers. A network typically has a set of base stations distributed across a geographical service area. A mobile device accesses the network at any given time through a wireless communication link with one of the base stations. As the device moves through the service area, however, the network must handoff the communication link to other base stations so that the service is not interrupted or degraded. Techniques related to handoff in wireless communication system are disclosed in U.S. Pat. No. 5,920,549 to Bruckert et al., U.S. Pat. No. 5,901,354 to Menich et al., U.S. Pat. No. 5,754,945 to Lin et al., and U.S. Pat. No. 5,657,487 to Doner. These patents are incorporated herein by reference in their entirety for all purposes.

[0004] In current wireless networks, the decision to execute a handoff to a nearby base station is typically performed by the network, often using received strength signal indicator (RSSI) information provided by the mobile devices to the base stations. The information provided from a mobile device typically includes values indicating the strengths of signals received at the mobile from several nearby base stations. The communication network then makes a handoff decision according to a hysteresis rule. If the signal power from a nearby base station exceeds by a predetermined hysteresis level that of the base station currently providing the link, then a handoff to the nearby base station is executed. This technique, however, is problematic due to variations in the signal power level that are not correlated to distance from the base station. Because the signal from the mobile device scatters from nearby objects, the signal power experiences variations as the mobile device moves over small spatial scales (i.e., distances on the order of a wavelength). In addition, there are signal power variations as the mobile device moves over larger spatial scales (i.e., distances on the order of building sizes). These variations are caused by shadowing of the signal by local obstructions such as buildings, hills, and the like. In order to mitigate the effect of small scale variations, conventional handoff techniques compute an average of signal samples collected as the mobile device moves over a few wavelengths, and then applies the hysteresis rule to this average signal. This technique, however, only mitigates small scale variations. The average signal still experiences considerable fluctuation due to shadow fading on larger scales. As a result, this local shadow fading often causes inappropriate and unnecessary handoffs that deteriorate the wireless system performance and degrade the quality of service to the mobile device.

[0005] The present inventors have discovered a novel technique for significantly reducing unnecessary and inappropriate handoffs in cellular communication networks, thereby improving service and system performance. Because shadow fading is repeatable due to its strong correlation to the spatial location of a mobile device in its environment, pattern recognition techniques can be used to identify shadow fading and prevent unwanted handoffs that would otherwise have been executed using conventional techniques. More specifically, the averaged signals of mobile devices that have followed the same geographical paths will have similar shadow fading properties. According to the technique of the invention, these averaged signals are stored along with information regarding handoff decisions. The signal from a mobile device that is traveling along a known path can then be compared with the stored signal to determine if a handoff is actually necessary. The inventors provide a technique to recognize these patterns in the signals from mobile devices using probabilistic neural networks, as well as a handoff method using this pattern recognition technique.

[0006] In one aspect of the invention, a method is provided for determining handoff in a wireless communication system comprising a plurality of base stations and a mobile device. The method comprises: (a) collecting at nearby base stations signal power level information representative of signal strength between the mobile device and the nearby base stations; (b) associating the power level information with predetermined power level patterns to produce an output representing measures of similarity between the power level information and the predetermined power level patterns; (c) selecting a sequence of matching patterns from the set of predetermined power level patterns, wherein each of the matching patterns has a maximal degree of similarity with a portion of the power level information; and (d) analyzing a sequence of sets of allowable base stations to determine an appropriate base station for handoff, wherein the sequence of sets of allowable base stations correspond to the sequence of matching patterns.

[0007] Preferably, the predetermined power level patterns are correlated with shadow fading in a propagation environment of the mobile device. More specifically, the signal power level information preferably comprises a sequence of average signal power level estimates corresponding to mobile device positions separated by less than a correlation length of a log-normal shadow fading in a propagation environment of the mobile device. Each of the average signal power level estimates is preferably an average of signal power level samples corresponding to mobile device positions separated by less than a wavelength of a carrier signal of the communication system. The signal power level samples are preferably measured at times selected such that the samples correspond to mobile device positions separated by approximately equal spatial intervals.

[0008] The step of associating the power level information with the predetermined power level patterns preferably comprises applying the power level information to a probabilistic neural network whose weights are equal to training vectors derived from training power level information. Other types of neural networks or learning machines, however, may also be used. The predetermined power level patterns preferably have a predetermined sequence corresponding to a sequence of patterns observed in a training session. The step of analyzing the sequence of sets of allowable base stations to determine the appropriate base station for handoff preferably comprises selecting a base station that is an allowable base station for a subsequent pattern in a predetermined sequence of power level patterns. In a preferred embodiment, the method also comprises determining a degree to which adjacent patterns in the sequence of matching patterns meet an adjacent pattern matching tolerance value.

[0009] FIG. 1 is a schematic diagram of a probabilistic neural network used in a preferred embodiment of the handoff determination technique of the invention.

[0010] FIG. 2 is a graph of the minimum value of the difference between the signal mean {overscore (s)}_kand the receiver threshold S_Tas a function of the probability of failure P_F,max.

[0011] FIG. 3 illustrates a possible geometry of base stations and two possible paths of a mobile device for the case of four nearby base stations.

[0012] FIGS. 4 and 5 are graphs of the difference between the minimum output signal and the required signal level versus the number of handoffs for paths 1 and 2, respectively.

[0013] FIG. 6 is a diagram illustrating a few segments of a path followed by a mobile device, together with corresponding signal samples and their groupings into pattern windows.

[0014] A preferred embodiment of the techniques of the present invention will now be described in the context of a typical cellular communications environment. Those skilled in the art, however, will recognize that the central ideas of the invention are not limited to many of the details below.

[0015] Wireless Propagation Model

[0016] An understanding of the techniques of the present invention is facilitated by the following description of a wireless propagation model.

[0017] The small scale signal variation considered in the present description is assumed to be caused by the sum of several waves with random phases and amplitudes with no direct path between the base station and the mobile device. Under these conditions, the envelope r_kof the received signal at a distance d_kfrom the base station to the mobile device can be modeled as a Rayleigh-distributed random variable with probability density function f(r_k) given by $\begin{array}{cc}f\left(r_k\right)=\frac{r_k}{p_k}\exp \left(-\frac{r_k^2}{2p_k}\right),r_k\ge 0,& \left(1\right)\end{array}$

[0018] where p_kis a parameter of the density function, and k is a discrete time index. The mean of the Rayleigh distribution is E[r_k]=(πp_k/2)^1/2. The small scale variation which can be modeled by equation (1) is called “Rayleigh fading”. The variation of received signal power with distance is modeled as 1/d_kⁿ, with n ranging from 2 to 6, depending on the degree of multipath in the environment. (In an open plain n=2,in an average suburban environment n=4,and in a dense urban environment n=6.)

[0019] The shadow fading is modeled as a correlated log-normal random process as is well-known in the art, e.g., as described in D. C. Cox, “Universal Digital Portable Radio Communication,” Proceedings of the IEEE, vol. 75, no. 4,pp. 436-477, 1987, and in M. Gudmundson, “Correlation Model for Shadow Fading in Mobile Radio Systems,” Electron. Lett., vol. 27, no. 23, pp. 2145-2146, 1991. Let R_L(d) denote the autocorrelation function of the shadow fading random process L(d), where d is the distance from the mobile to the base and L(d), measured in decibels (dB), is a normally distributed random process. The autocorrelation, R_L(d), is given by: $\begin{array}{cc}{R}_L\left(d\right)={\sigma }_L^2\exp \left(-\frac{\left|d\right|}{d_0}\right),& \left(2\right)\end{array}$

[0020] where σ_L²and d₀are, respectively, the variance and correlation length of L(d). From equation (2), the power spectrum S_L(V) of L(d) is given by: $\begin{array}{cc}{S}_L\left(v\right)=\frac{2d_0{\sigma }_L^2}{1+{\left(2\pi \nu d_0\right)}^2},& \left(3\right)\end{array}$

[0021] where v is the spatial frequency. For a total distance travelled, D, the shadow fading process can be shown to be: $\begin{array}{cc}L\left(d\right)=\sum _{j=-J}^J\sqrt{\frac{2}{\mathrm{BD}}{S}_L\left(\frac{j}{D}\right)}\cos \left(\frac{2\pi \mathrm{jd}}{D}+\phi \right),& \left(4\right)\end{array}$

[0022] where $\begin{array}{cc}B=\frac{1}{{\sigma }_L^{2}D}\sum _{j=-J}^J{S}_L\left(\frac{j}{D}\right),& \left(5\right)\\ J=D{\nu }_m,& \left(6\right)\end{array}$

[0023] where v_mis the maximum spatial frequency taken into account, and φ_j, is an independent, identically distributed uniform random process in [0,2 π). The process L(d) is sampled at d=d_kto obtain L_k. (In typical applications of the present invention, the sampling period is preferably between 5 and 35 ms/sample, and can be constant or variable.)

[0024] In the absence of Rayleigh fading, the signal is expressed in dB by

[0025] m_k=C₀−10nlog₁₀d_k+L_k,  (7)

[0026] where C₀is a constant which includes the power transmitted, antenna parameters, and carrier frequency. Thus, the received signal, including Rayleigh fading (in dB), is given by:

[0027] s_k=20log₁₀r_k,  (8)

[0028] where r_kis a Rayleigh distributed random variable with parameter p_k=(1/2)^(m/10). The mean and variance of s_kare given, respectively, by: $\begin{array}{cc}{\stackrel{\_}{s}}_k=10{\log }_{10}\left(2p_k\right)-\frac{10\gamma }{\ln 10}=m_k-\frac{10\gamma }{\ln 10},& \left(9\right)\\ {\sigma }_{s_k}^2=\frac{50{\pi }^2}{3{\left(\ln 10\right)}^2},& \left(10\right)\end{array}$

[0029] where γ≈0.577216 is Euler's Gamma. The signal is non-stationary since {overscore (s)}_kis not constant and varies according to equation (9). The model presented here is used below for performance evaluation of the hysteresis rule and the pattern recognition handoff algorithm.

[0030] Pattern Recognition Using Probabilistic Neural Networks

[0031] In the preferred embodiment, a probabilistic neural network (PNN), whose structure is shown in FIG. 1, is used in the handoff algorithm. This network, or an equivalent learning machine, is implemented in a signal processor of the communication system. Signal processors are used in wireless communication systems that use conventional handoff techniques. The implementation of the present invention is similar to conventional systems with the primary exception of the operation and structure of the signal processor used for determining handoff. Thus, the following description will focus primarily on these differences, as those skilled in the art will already understand and appreciate the conventional architecture and operation of a wireless communication system having handoff management capabilities.

[0032] The following notations will be used to describe the PNN shown in FIG. 1. Let X=[x₁,x₂, . . . , x_n]^tdenote a test vector to be classified (i.e., matched with a pattern). Such a vector is a portion of signal power information collected at several base stations. As illustrated in the figure, the vector is applied to the PNN from the left and its components are fed to the various neurons, labeled W₁, W₂, W₃, W₄, . . . , W_P−3, W_P−2, W_P−1, W_p. Let X_p=[x_p,1,x_p,2, . . . , x_p,N]^Trepresent the p-th training vector (i.e., pattern), let c_pbe the pattern label (i.e., class) of X_p, and let W_p=[w_p,1,w_p,2, . . . , w_p,N]^Tbe the weight vector of the p-th neuron in the PNN. The number N is a design parameter that preferably has an integral value of at least 50. If computational complexity is a consideration, N is preferably less than 300. The output y_pof the p-th neuron is given by: $\begin{array}{cc}{y}_p=\exp \left(-\frac{{\Vert X-W_p\Vert }^2}{{\sigma }_N^2}\right),& \left(11\right)\end{array}$

[0033] where σ_N²is a “smoothing parameter” of the PNN. Those skilled in the art will appreciate that other substantially similar formulas can be substituted for relation (11), provided they measure a degree of similarity between X and W_p. Note that if X=W_p, then y_p=1, and as the difference between X and W_pincreases, y_papproaches 0.

[0034] Let P be the number of training vectors or neurons (p=1,2, . . . , P) and C be the number of classes of patterns (c_p=1, 2, . . . , C). A class of patterns corresponds to a recognizable segment of a path of the mobile device. These segments are typically 50 m to 200 m portions of a path followed by the mobile device. In a preferred implementation P=C since only one pattern is required to recognize each segment of a path. The total number C of pattern classes is equal to the total number of recognizable segments in all the paths that the system has been trained to recognize. The value of C is preferably in the range from 10 to 1000 or more. In alternate embodiments P>C since multiple training vectors can be used to recognize each segment of a path (e.g., corresponding to different seasons or other environmental changes). The weights of the neurons are set equal to the training vectors,

[0035] W_p=X_p, p=1,2, . . . , P.  (12)

[0036] Y_α={y_p:c_p=α}, α=1,2, . . . , C,  (13)

[0037] $\begin{array}{cc}{z}_{\alpha }=\sum _{y_p\in Y_{\alpha }}{y}_p,\alpha =1,2,\dots ,C.& \left(14\right)\end{array}$

[0038] The set of outputs y_pof all neurons whose training patterns X_pbelong to class α is Y_α, and the sum of the elements of Y_α is z_α. The association of neurons with classes (i.e., the manner in which the neurons are “wired” to the adders in FIG. 1) is determined during training for use during real time pattern recognition. The test vector X is associated with class ĉ ; according to the following rule:

[0039] ĉ ;=arg max_αz_α.  (15)

[0040] The above classification rule assumes that the ratio of the number of training patterns for class α to the a priori probability of occurrence of class α is constant for all classes α=1,2, . . . , C.

[0041] As shown in FIG. 1, the outputs of the neurons are fed to a set of adders that produce the values z₁, Z₂, . . . , z_Cin accordance with equations (13) and (14) above. The processing block labeled “arg max_α” then determines the class ĉ ; of the vector X originally fed into the PNN.

[0042] Handoff Method Based on Recognition of Patterns in Signal Power

[0043] A handoff technique according to the present invention uses the following performance criterion: a handoff is executed to ensure that the probability of link failure, P_F, is less than or equal to a specified value P_F,max. The value for P_F,maxis preferably 0.02 or less, since higher failure rates result in degradation of the signal beyond the subjective tolerance of listeners. For this analysis, failure occurs if the received signal s_kfalls below the receiver threshold S_T, which is an intrinsic parameter of the receiver sensitivity. The design requirement is then:

[0044] P_F≡Pr{s_k<S_T}≦P_F,max.  (16)

[0045] From equations (1) and (8), we have: $\begin{array}{cc}{P}_F=\mathrm{Pr}\left\{r_k<{10}^{S_T/20}\right\}=1-\exp \left\{-\frac{{10}^{S_T/10}}{2p_k}\right\}.& \left(17\right)\end{array}$

[0046] An inequality for {overscore (s)}_kcan be derived using equations (9) and (16), with the result $\begin{array}{cc}{\stackrel{\_}{s}}_k\ge S_T-10{\log }_{10}\left[\ln \left(\frac{1}{1-P_{F,\max }}\right)\right]-\frac{10\gamma }{\ln 10}.& \left(18\right)\end{array}$

[0047] FIG. 2 plots the minimum value of {overscore (s)}_k−S_Tas a function of P_F,max. For P_F,max<<1, the following approximation to equation (18) can be used: $\begin{array}{cc}{\stackrel{\_}{s}}_k\ge S_T-10{\log }_{10}\left(P_{F,\max }\right)+\frac{5P_{F,\max }-10\gamma }{\ln 10}& \left(19\right)\end{array}$

[0048] In practice, {overscore (s)}_kis not easily measurable. An estimate ξ_kof {overscore (s)}_kis obtained by averaging the past M samples of s_k: $\begin{array}{cc}{\xi }_k=\frac{1}{M}\sum _{m=0}^{M-1}{s}_{k-m}.& \left(20\right)\end{array}$

[0049] The value of M is preferably 10 to 100. The specific value used is a design parameter that depends on the sampling rate and the correlation length of the log-normal shadowing. The samples s_kare measured at the mobile and preferably transmitted to the base station on the control channel for averaging at the base station. Alternatively, the mobile performs the averaging and transmits the estimates ξ_kto the base station. It is noted that conventional handset hardware contains received signal strength indicator (RSSI) circuitry that can be used to determine the value of the samples. In the preferred embodiment, the s_kvalues are obtained by sampling at equal spatial intervals. This assumption is equivalent to constant velocity of the mobile device and temporal sampling at regular intervals, or variable velocity and temporal sampling at correspondingly irregular intervals. Various known velocity estimation techniques can be used to estimate the mobile device velocity. Once the velocity is known, the sampling rate can be adjusted to ensure that the sampling corresponds to regular spatial intervals. Furthermore, the distance spanned by the samples s_k−(M−1), . . . , s_kis preferably less than ⅓ the correlation length do of the shadow fading process, and more preferably less than {fraction (1/10)} the correlation length. Moreover, the distance between adjacent samples is preferably at least λ/2, so that adjacent samples are approximately independent. Under these conditions, ξ_kis the sum of independent, identically distributed random variables. For sufficiently large M, the Central Limit Theorem states that ξ_kapproaches a Gaussian distributed random variable with mean {overscore (ξ)}_k={overscore (s)}_kand variance ρ_ξ²=σ_s²/M. For {overscore (s)}_kto satisfy equation (18) with 99% confidence, we require $\begin{array}{cc}{\xi }_k\ge S_T-10{\log }_{10}\left[\ln \left(\frac{1}{1-P_{F,\max }}\right)\right]-\frac{10\gamma }{\ln 10}+3{\sigma }_{{\xi }_k}+{\xi }_{\mathrm{margin}},& \left(21\right)\end{array}$

[0050] where ξ_marginis a margin of safety. This margin of safety is preferably in the range σ_ξto 3σ_ξ.

[0051] Using relation (21) on the averaged signal, a handoff technique based on pattern recognition is now described. A test vector is derived from a set of averaged signal samples from a few (2 to 4) of the nearby base stations B_i(i=1, . . . , i_max) having the strongest signals. Samples from each of these base stations are collected within a spatial window corresponding to a segment of the mobile device path, as shown in FIG. 6. Let N_wdenote the number of averaged samples collected for each base station along a segment, i.e., N_wis the sample length of the spatial window. The total number of samples in a pattern vector will thus be a multiple of N_w. A training run is made along each path to initialize the PNN for pattern recognition. The averaged signal samples from the nearby base stations are recorded for each path. Let ξⁱ=[ξⁱ₀, ξⁱ₁, . . . , ξⁱ_N−1]^Tand ξ^j=[ξ^j₀, ξ^j₁, . . . , ξ^j_N−1]^Tdenote the averaged samples from base B_iand base B_j, respectively, where the total number of samples recorded from each base station along a given path is N_max. The number of classes of patterns in the path is therefore

[0052] C=N_max/N_w.  (22)

[0053] For each distinct training class α_t, t=0,1, . . . , C−1, a training vector X_t^FWDis formed from all the samples within the corresponding spatial window, including samples from multiple nearby base stations. The class label α_trepresents the segment index, or also can represent the midpoint of the spatial window. The ordering of the samples in the training vector is arbitrary, provided it is a consistent ordering. Preferably, the ordering is sequential and corresponds to travel of the device in the same direction as the training run. A second training vector X₁^REVfor the same class α_tis formed by reversing the order of the samples of X_t^FWDto account for travel in the opposite direction of the training run. Thus, each class has two vectors that are dependent on each other, which is equivalent to one independent vector. Indeed, rather than storing both these vectors, the sample order can be reversed in real time during pattern matching, trading off memory for computation. Most pattern recognition applications require more than one training vector per class. However. in this handoff technique, one independent training vector per class has good results since the use of averaged signals together with class sequencing contributes to fewer miscalculations by the PNN.

[0054] Once the training vectors from the training run are stored, the patterns which correspond to handoff locations are determined as follows. The minimum sample, ξⁱ_t,min, corresponding to base B_iis determined for each class α_t(i.e., for each segment of the path):

[0055] ξⁱ_t,min=min(ξⁱ_tN,ξⁱ_tN+1, . . . , ξⁱ_(t+1)N−1), i=1,2, . . . , i_max.  (23)

[0056] Base B_iis an acceptable serving base station for class α_tif ξⁱ_t,minsatisfies relation (21).

[0057] The method of determining handoff for a mobile device is described now. As a mobile device travels along the path, a vector of the averaged received signal samples from various base stations is fed as input to the PNN, with the vectors corresponding to segments of the path. The direction of travel by the mobile device can be deduced by observing the order of the outputs of the PNN, and a future need for handoff can be anticipated. To account for misclassifications (due to lack of pattern window synchronization, for example), a tolerance T_Mcan be set for the difference between PNN outputs corresponding to the adjacent pattern windows. The value for T_Mis preferably in the range from 1 to 10, but is more preferably in the range from 2 to 4. Let ĉ ;_kdenote the class output for the k-th pattern window, β_kdenote the corresponding set of allowable serving base stations (determined during training), and B_kεβ_kdenote the serving base station. Also let Δα be the distance between adjacent training class labels. This distance Δα is typically 1 when α represents a class pattern index, and is typically some physical distance when α represents a location of a segment along a path. Conditions (24)-(27) are computed for the determination of handoff:

[0058] ∥ĉ ;_k−ĉ ;_k−1∥≦(Δα)T_M,  (24)

[0059] {B_k}∩β_k+j=Ø, for some jε{1,2, . . . , T_M},  (25)

[0060] $\begin{array}{cc}\bigcap _{i=1}^{l+!}{\beta }_{k+i}=\varnothing ,\mathrm{for}l>0,& \left(26\right)\\ \bigcap _{i=1}^l{\beta }_{k+i}\ne \varnothing ,& \left(27\right)\end{array}$

[0061] Condition (24) is satisfied when the recognized classes for adjacent windows are “close” within a tolerance T_M. If this condition is not satisfied, a window synchronization procedure shifts the alignment of the window frame with respect to the received samples to scan for proper synchronization. Condition (25) is satisfied when the current base station is not an allowable base station at some point within the tolerance T_M, indicating that a handoff is needed in the near future. Conditions (26) and (27) are used to determine a base station to which handoff is to be performed. These conditions are satisfied when some base station is allowable for a number of windows into the future, but is no longer allowable past this critical number of windows. If conditions (24)-(27) are satisfied, a handoff is made to an element of $\bigcap _{i=1}^l{\beta }_{k+i}.$

[0062] Simulations of the technique described above in the case of typical implementations illustrate the performance advantages of the present invention over conventional techniques that use the hysteresis rule. The simulations were performed for the cases of four and five nearby base stations. For simplicity, a mobile device is assumed in the simulation to move in a straight line at constant velocity. In this case, sampling at equal time intervals corresponds to sampling at equal spatial intervals.

[0063] Let sⁱ_kand s^j_kdenote the signals received at the mobile from bases B_iand B_j, respectively. Assuming the mobile device is initially communicating with base B_i, the hysteresis rule is as follows: a handoff from base B_ito base B_jis performed if max{s^t_k:t=1, . . . , i_max}=s^j_kand sⁱ_k−sⁱ_k>H. The hysteresis level H is specified in dB. Simulations are performed for several environments. FIG. 3 illustrates the geometry of the base stations and two paths of the mobile device for the case of four nearby base stations. The propagation model discussed earlier is used with independent log-normal shadow fading for a given environment, path, and base station. Table 1 below indicates the parameters used in the simulations, which also represent preferred values for use in a preferred embodiment corresponding to the simulation environment.

[0064] Simulations were performed for 50 different environments for four paths of the mobile device for each base station geometry. Only one training run was used for each environment and path. In reality, the test runs might start and end at different locations with respect to the training runs. This aspect is accounted for by aligning the test patterns with the training patterns to minimize misclassifications. For each environment, four paths of the mobile device and five test runs per path were considered. Thus, the total number of test runs is 1000 for each case of four and five base stations.

[0065] FIGS. 4 and 5 are plots of the difference between the minimum output signal for the test runs and the required signal level (obtained from relation (21)) versus the number of handoffs for paths 1 and 2. Each figure corresponds to a particular path for the case of four base stations. Performance results for various other paths of the mobile device and for the case of five nearby base stations are qualitatively similar to the results presented here. Each data point represents an average over five runs for a given environment. For clarity, the region from 0 to 5 handoffs has been magnified. The plots compare the handoff technique using a probabilistic neural network (PNN) with the conventional hysteresis rule using hysteresis levels of H=5 dB and H=10 dB. The hysteresis level H cannot be increased indefinitely to yield lower handoffs without resulting in an increased failure rate and possible handoffs to incorrect base stations. Out of a total of 2000 test runs, use of the PNN resulted in 99% of the runs having minimum signal levels above the required signal level for 2% probability of failure. Furthermore, around 97% of the runs resulted in accurate classification by the PNN. It is important to note that only one training run was used for each environment and path. The misclassification rate can be reduced if more training runs are used, e.g., adding more training vectors for each pattern class (i.e., segment of the path). In an alternate embodiment, additional training vectors can be added to account for seasonal changes in the propagation environment. For a given probability of failure, the pattern recognition based handoff technique performs fewer handoffs while the hysteresis rule executes several unnecessary handoffs. In some instances, the pattern recognition based handoff technique executes no handoffs because the signal is sufficiently strong to satisfy relation (21). In reality, one would expect a handoff as the mobile approaches another base station with a stronger signal. This performance can be achieved by using a large overlay hysteresis. Thus, in a preferred embodiment, the pattern recognition technique described above is supplemented by an overlay hysteresis, as follows. If the pattern recognition technique does not result in a handoff at a given point in time, then the power level of the minimum signal in the current window from the various base stations are compared. If the signal from another allowable base station has a power level at least 20 dB larger than that from the current base station, then handoff to the other base station is executed.

[0066] The handoff technique based on pattern recognition can be extended to other possible performance criteria. For example, the acceptable base stations can be ranked and this information can be used to improve handoff to the best base station. The method can be extended to consider mobile devices changing paths at intersections with variable speeds. For example, a velocity estimator can be used in a preprocessing stage, and velocity information can then be used to appropriately adjust the sampling rate. In addition, unsupervised learning methods can be applied to eliminate the necessity of training runs. For example, real time data can be used for training by correlating handoff information using conventional hysteresis methods with measured signal samples to identify new paths. Once the paths are identified, the pattern recognition technique can be used to anticipate and initiate handoff. It will be appreciated by those skilled in the art that the concepts and techniques of the present invention do not need to be implemented using probabilistic neural networks. Other implementations that are well known in the art can also be used to achieve the same effect. For example, various feature extractors, pattern classifiers, learning machines, and neural network architectures can be used. Thus, the spirit and scope of the present invention should not be limited to the specific details of the embodiments described above.