WO2008153798A2

WO2008153798A2 - Channel state inference/prediction using observable variables

Info

Publication number: WO2008153798A2
Application number: PCT/US2008/006745
Authority: WO
Inventors: Shirish S. Karande; Syed Ali Khayam; Yongju Cho; Hayder Radha; Jaegon Kim; Jin-Woo Hong
Original assignee: Board Of Trustees Of Michigan State University; Electronics And Telecommunications Research Institute
Priority date: 2007-05-30
Filing date: 2008-05-28
Publication date: 2008-12-18
Also published as: KR20080106088A; KR101006394B1; WO2008153798A3

Abstract

Feasibility of utilizing observable variables in channel state inference and channel state prediction was investigated. For example, measurements of signal to noise ratio and background traffic can be used to improve the accuracy of channel state inference. A link-invariant model for estimating bit error rate of a data packet was derived using freely observable parameters, such as signal to noise ratio and/or background noise. Prediction can also be achieved by considering the estimate for a current data packet as a prediction for a future data packet or by utilizing a link-specific model that captures the correlation in bit error rate of temporally adjacent packets.

Description

CHANNEL STATE INFERENCE/PREDICTION USING OBSERVABLE VARIABLES

FIELD [0001] The present disclosure relates generally to channel state inference and/or prediction using observable variables.

BACKGROUND [0002] Unlike the traditional wired Internet based communication, the number of packet drops due to bit errors can be substantial in wireless networks. Bandwidth hungry multimedia systems can be adversely affected by such packet drops. Therefore many recent multimedia related studies have recommended the development of cross-layer protocols that do not discard partially damaged packets. Relay of corrupted packets to higher layers can lead to a significant amount of improvement in the video throughput. However, the efficacy of information recovery from a corrupted packet is a function of the bit error rate (BER) in the packet. Hence, accurate channel state information (CSI) or channel state prediction (CSP) can provide substantial capacity gains.

[0003] Utility of channel awareness in improving capacity is a well researched area and has been demonstrated by many theoretical/simulation based studies (particularly the ones that focus on the physical layer). However unlike the physical layer, the channel observed at the link-layer (MAC) is discrete, and each individual bit does not have a signal strength associated with it. This makes the task of practically providing CSI about the corruption levels in a packet significantly harder. Therefore, it is desirable to identify and develop practical mechanisms that can provide CSI and CSP at the link-layer.

[0004] The statements in this section merely provide background information related to the present disclosure and may not constitute prior art.

SUMMARY

[0005] Feasibility of utilizing observable variables in channel state inference and channel state prediction was investigated. For example, measurements of signal to noise ratio and background traffic can be used to improve the accuracy of channel state inference. A link-invariant model for estimating bit error rate of a data packet was derived using such freely observable parameters, such as signal to noise ratio and/or background noise. Prediction can then be achieved by considering the estimate for a current data packet as a prediction for a future data packet or by utilizing a link-specific model that captures the correlation in bit error rate of temporally adjacent packets.

[0006] Further areas of applicability will become apparent from the description provided herein. It should be understood that the description and specific examples are intended for purposes of illustration only and are not intended to limit the scope of the present disclosure.

DRAWINGS

[0007] Figures 1A and 1 B are diagrams of exemplary topologies for wireless trace collection;

[0008] Figures 2A-2F are charts illustrating average statistics deduced from error traces;

[0009] Figure 3 is a graph showing the average value of BER over corrupted packets as a function of SSR; [0010] Figure 4 is a flowchart illustrating a method for deriving models for estimating BER of a data packet;

[0011] Figure 5A and 5B are charts showing the concentration gains for two exemplary traces;

[0012] Figure 6 is a chart illustrating the temporal correlation in BER between data packets;

[0013] Figure 7 is a flowchart illustrating a method for predicting the BER of a data packet; and

[0014] Figures 8A and 8B are graphs showing the concentration loss for two exemplary traces. [0015] The drawings described herein are for illustration purposes only and are not intended to limit the scope of the present disclosure in any way. DETAILED DESCRIPTION

[0016] A measurement based study of 802.11b wireless LANs is presented that analyzes the utility of observable variables in channel state inference and prediction. For this study, consider two wireless setups for collecting error traces as shown in Figures 1A and 1 B. In Figure 1A, five wireless receivers were used to simultaneously collect error traces on an 802.1 1 b WLAN. One receiver was placed within clear line-of-sight (LoS) of the access point (AP), while the remaining four receivers were placed at different locations in a room across the hallway (referred to herein as setup A). In Figure 1 B, six receivers were used to simultaneously collect error traces (referred to herein as setup B). Three receivers were placed in a room across the hallway, while three receivers were placed (at an extreme edge of the network) in a room 100 feet down the hallway. A wired sender was used to send multicast packets with a predetermined payload on the wireless LAN; multicasting disabled MAC layer retransmissions. Each experiment comprised of one million packets with a payload of 1,000 bytes each. At the physical layer, the auto rate selection feature of the AP was disabled and for each experiment the AP was forced to transmit at a fixed data rate. Each trace collection experiment was repeated for 5.5 and 11 Mbps physical layer (PHY) data rates. For each PHY data rate and for each SETUP, we collected traces for two distinct packet transmission rates. The transmission rate is controlled by adjusting the time interval * between packets. In setup A we collected traces at 500 Kbps and 1024 Kbps, while in Setup B we collected traces at 750 Kbps and 900 Kbps. For ease of notation, we prefer to label the traces by their PHY data rate and single number. Table I describes the numbering we shall use for the remainder of the work.

Table I Trace numbering key for 5.5Mbps and 11 Mbps

While the following description is provided with reference to 802.11 b networks, it is readily understood that the concepts disclosed herein are broadly applicable to wireless networks.

[0017] The receivers' MAC layer device drivers were modified to pass corrupted packets to higher layers. To capture packets at high transmission rates, packet dissectors were implemented inside the device drivers. These packet dissectors ensured that only packets pertinent to our wireless experiment are processed, while all other packets are simply dropped. In addition to a packet's header and payload information, for each packet three additional parameters were logged at the receivers:

• Background traffic (B): A four byte number representing the total number of background packets observed between two trace packets;

• Signal strength (S) for the received packet: A one byte number representing the signal strength in dBm; • Silence Value (N) for the received packet: A one byte number which can be said to be representing the noise + interference strength in dBm. We use the following defining equations to associate a a signal to silence ratio (SSR) and background traffic (BT) intensity, p , with each packet:

It is important to note that in typical 802.11 b wireless receivers, the silence value N is measured just before the packet reception starts and the signal strength is measured for only the first few μs ( -10). Thus the SSR indication associated with each packet is only an approximate measure of the link-quality experienced by a packet transmission. Similarly, also note that in measuring B we do not differentiate between packets on the basis of their size, signal strength, PHY rate etc. Thus our measurement of the background traffic intensity, p , is relatively coarse and can be improved upon. However, even such coarse measurements are sufficient to provide predictive utility. Other techniques for deriving these parameters are also within the scope of this disclosure. Moreover, it is envisioned that other types of observable parameters may be used. [0018] Figures 2A-2F show some of a set of average statistics deduced from the error traces we consider in this work. These average statistics should provide a good representation of the wireless environment in which we have conducted our experiments. Figure 2A and 2D show δ the proportion of packets that were corrupted; Figure 2B and 2E show the average value of SSR; and Figure 2C and 2F show the average value of BT intensity p. It can be seen that trace 3 and 20 which correspond to the LOS client have a very good link quality and rarely see any packet corruptions. Thus we often exclude these traces from our analysis. The value of p was less than 50 packets/s for most traces but for certain traces the intensity can be well above 200 packets/s. Comparing the plots for 5.5MBps and 11 MBps it can be seen that the long-term average value of SSR for a specific link does not vary significantly, however the same cannot be said about the BT. Additionally, note that as expected the packet corruption ratio is lower when the PHY data rate is 5.5Mbps as compared with the 11 Mbps traces.

[0019] Next, we shall investigate the utility of observable variables as side-information in CSI. By CSI, we specifically refer to a problem where we want to estimate the BER in a packet that has already been received. Accuracy of such estimates plays an important role in soft-decoding algorithms and can also be important for a variety of other reactive protocols.

[0020] Figure 3 shows the average value of BER (θ ), over all the corrupted packets as a function of SSR, for the cases when there is no BT and when there is heavy BT. Thus it can be clearly seen that the BER in a packet is correlated with the BT and SSR. Moreover the relationship of SSR with BER can vary in presence of BT. Thus it may be feasible to utilize SSR and p as side-information and there may be a necessity to use them jointly for robust CSI. [0021] A received packet ^~x is said to be of type τ_θ if the (sample)

BER in the packet is θ . We can use such a definition of type, purely because * is made up of binary symbols and the probability distribution on these symbols is completely defined by the parameters . All the discussion in this section is strictly for the case of binary symbols and often we shall use θ to actually represent the probability distribution on the binary symbols. [0022] One objective is to infer or predict the type of a packet as accurately as possible. Additionally we assume that the side-information Z is able to accurately determine the type T₀, i.e. the packets with no errors (0 = 0).

Hence we concentrate all our analysis only on packets with type τ_θ ≠ τ₀. Therefore we can represent the cost of estimating/predicting the type of a packet as θ by

where DΪølløj is the Kullback-Leibler divergence measure. Thus f(θ,θ\ can be interpreted as a "loss in capacity" on account of assuming that the binary process ^~x is governed by the Bernoulli parameter θ when it is actually governed by θ . This "capacity loss" interpretation can occur, for example, if additional information needs to be transmitted to compensate for the distance o(ø|ø) - Further, it is important to note that the "loss in capacity" takes place

regardless if the estimate θ is pessimistic or optimistic relative to the true BER value (type) e . The performance measures we choose are influenced by the discussion on Method of Types presented in "Elements of Informational Theory" by T. M. Cover and J.A. Thomas Wiley Series in Telecommunications which may be referred to for additional background information.

[0023] A data set A , which is a set of packets, enforces a probability distribution _/>_A (-) on the types τ_β or on the parameter θ . _/>_A (ø) represents the frequency with which a packet of type τ_β is observed in a data set A . Our approach is based on choosing a representative type r. for a set A so that the average cost of mis-representing members of A by type T- is minimized. Thus from a given set we determine our model parameter as

θ_Λ = *rgmin(E_pΛ [f {θ,θ')]) (2)

where £_PA [/(Ø,Ø' )] is the Expected KL-Divergence (EKLD). Since f (θ,θ') represents the loss in Shannon's information (capacity), £^ [/(0,0^*)] can be interpreted as the loss in the ergodic information (capacity). Note that E \f {θ,θ') \ measures the concentration of the types in set A around the type T- .

[0024] Since in this case we are just operating on a binary symbol space the estimate representative Type (i.e. the model parameter θ_A ), in accordance to (2) can be achieved simplistically. The following lemma shows that parameter estimate can be achieved by just taking the expectation in the data set.

Lemma 1 : lf /(ø,ø^*) = D(ø|<9^*) , A s.t v le A ø(x) ≠o and θ is given by (2) then

Proof: The proof follows by solving -?-_rE_PA \f (θ,θ')] = o .

In the above discussion, our model was completely defined by a single parameter θ ; however the data set we consider in this work is labeled. In particular, each ^~x is labeled by SSR and p . Thus we can obtain a set- decomposition of the training set A utilizing one or both the labels. Let us denote these subsets by A_SSR , or A_P , or A_{(SSR p)} , depending on the particular label(s) used. For each subset so obtained, we can determine a distinct parameter, which can be represented as Θ(SSR) , or θ(p) or Θ(SSR,P) . These individual parameters are obtained by just taking the expectation in the subset e.g: Θ(SSR,P) = E_PA [Θ] . [0025] With reference to Figure 4, this method for deriving models for estimating BER of a data packet may be summarized as follows. Models are derived from a plurality of data packets which serve as training data and are indicative of the data link over which data packets traverse. Each data packet in the training data is labeled at 41 with at least one observable parameter, such as SSR and/or BT, and actual bit error rate for the packet. It is understood that this technique may be extended to other types of observable parameters associated with a data packet.

[0026] Data packets having similar observable parameters are grouped together at 42 to form groups of data packets. In an exemplary approach, groups are empirically formed. However, it is envisioned that other techniques for grouping the data packet may be employed.

[0027] A cost function for estimating bit error rate is defined at 43 in the manner described above. For each group of data packets, the cost function is then minimized at 44, thereby determining an estimated bit error rate for data packets having the observable parameter associated with the given group of data packets. When a data packet having an unknown BER is received, a corresponding model is selected based on its observable parameters. This technique may be used to develop models for estimating other types of unobservable parameters of a data packet from freely observable parameters associated with the data packet.

[0028] The models obtained in this manner may be tested on various traces. Each trace Λ is represented by a vector time series {<?,, Z₁, SSR,, p,)_{l=l N} where θ,, z,, SSR₁, p, are obtained from the f^h packet in the trace. We say z = o if the packet is error free and z = i otherwise. As has been explained before we focus our analysis entirely on packets for which z = \ . N simply represents the total number of packets in each trace. We use the correlation models to obtain an estimate series {00)},_{=1 Λ}, . The utility of the model M, for a particular trace, in estimating the type of a packet is obtained as

where N₁ represents the number of packets with z = i in trace Λ . Depending on the model M being employed we can utilize SSR,, p, as side-information to obtain the estimate S, . In the absence of any side-information S₁ =S where S is the average statistic, but when side-information is utilized we can write the estimate as S₁ = S(SSR₁) , S₁ =S(P₁) , S,(I) = S(SSR,,P, ) . Since we are concentrating all our analysis on corrupted packets it is implicit that Z is always being used as side-information. Thus in the remainder when we discuss about the presence or absence of side-information we are specifically referring to the use of SSR and BT in addition to Z. [0029] Thus, we consider 4 models, which we refer to as Average, SSR_aware, BT_aware, SSR+BT_aware depending on the side-information they utilize. In this section, the model M ≡ {Ave., SSR, p, ssR+p) . For each PHY data rate we consider a distinct training set and this set is composed of all the 22 traces collected at the data rate. We test the performance of models so obtained on each of the traces. Depending on the model being deployed, we get 4 measures for each traces which can be denoted by EKLD_ΛVC [A] , EKLD_SSR [A] ,

EKLD _p [Λ] and EKLD_SSK+P [Λ] . We are interested in quantifying the improvement obtained by utilizing side-information in CSI. For this purpose we define concentration gain as:

concentration gain = 10 • log,₀ *"j ^J (5)

I EKLD_M [A] J where M can be any of {Ave., SSR, p, SSR + P) .

[0030] Figure 5A and 5B show the concentration gains for 5.5Mbps and 11 Mbps, respectively. It can be clearly seen that both BT and SSR can provide concentration gains. Note a gain of 3dB implies an improvement in concentration by a factor of 2. Thus for 14/22 traces collected at 11 Mbps and for 10/22 traces collected at 5.5MBps, utilizing SSR as side-information can improve the accuracy of estimating BER by at least a factor of 2. It can be clearly seen that on some traces the improvement is in excess of 24 dB which corresponds to an improvement in concentration by a factor greater than 250. Thus clearly the improvement in CSI by just utilizing SSR, despite the presence of BT, can be significant in many practical scenarios. As compared to SSR, the gains provided by BT are modest. However, there still exist a few traces, e.g. 6, 1 1 , 14, 19, 21 at 11 Mbps and 11 , 12, 16 at 5.5Mbps where the gain is close to or above 3dB. In certain cases it is possible to combine SSR and BT, to jointly use them as side-information, to achieve further gains. For example, see traces 11 and 14 for 11 Mbps, and traces 11 , 12 and 16 for 5.5Mbps.

[0031] An accurate estimate for BER of a data packet has many uses. One exemplary use is to improve the error recovery process for corrupt data packets received in a wireless communication system. Briefly, the BER is estimated for each individual data packet received at a receiver. The BER is estimated at a data link layer of the receiver or some other layer receiver below an application layer as defined by an Open System Interconnection (OSI) model. The BER for each data packet is then passed to an application layer of the receiver and an error recovery operation is performed in relation to a corrupt data packet using the BER associated with the corrupt data packet. More specifically, the BER is translated to a probability that a given bit in the data packet is in error and this probability is in turn used to decode each bit within the corrupt data packet. Details regarding how the BER is captured and used to perform an error recovery operation may be found in U.S. Patent Application No. 1 1/725,242 filed on March 16, 2007 and entitled "Method to Utilize Physical Layer Channel State Information to Improve Video Quality" which is incorporated herein by reference. Other uses for the estimated BER are also within the broader aspects of this disclosure.

[0032] An accurate estimate for BER also proves beneficial in channel state prediction (CSP). The problem of CSP specifically refers to predicting the state of packet that has not been received as yet and thus we have no information about the SSR and p values associated with such a packet too.

Hence it should be evident that the CSP problem considered below is distinct from the CSI considered above. The CSP can be employed or communicated to the transmitter with the help of feedback. Thus robust CSP can be used for controlling the rate of source and channel codes.

[0033] Many research studies have observed that temporal correlations can be observed in the channel states in 802.11 b wireless networks.

Figure 3 shows the correlation coefficients calculated on the basis of the BER process. The coefficient can be calculated simply as r_ ^gtø^g _|τ,]-E[fl,]E[έU Vvartø]-vartø₊₁] where the mean and variance are just the sample mean and sample variances calculated on the traces. Figure 6 clearly exhibits the existence of temporal correlation (at times significant). [0034] Temporal correlation in BER between two adjacent packets can be be taken advantage of to predict the BER in the future packet. This can be done in a simplistic manner by utilizing the BER in the current packet as an estimate for the BER in the next packet:

L = O₁ (6)

However, an improved performance can be obtained by developing a link- specific model that takes advantage of the temporal variations specific to the link. This can be done by empirically evaluating the conditional probability distribution P_A (Θ_I+1

) , by just observing the frequency with which various Type pairs (τ_βι,τ_θ^ ) occur in trace Λ . On the basis of this distribution we get an improved prediction as:

L = sΛθ_i) = E_pΛθM) [θ^} (7)

In an exemplary embodiment, the model may be a Markov model. Other types of statistical models, such as Hidden Markow models, hierarchical Markov models or multifractal models, are contemplated by this disclosure.

[0035] Models developed above for estimating BER cannot be employed for prediction in many practical situations because θ(ή is not observable wheneverz(*) ≠ i . In such circumstances we can utilize the side- information based CSI mechanism developed in the section above to estimates, . Due to brevity we shall focus only on utilizing SSR as side-information. Thus the predictor described by (6) can be realized as

Θ_I+1 = Θ(SSR, ) (8)

We realize the predictor given by equation (7) by estimating Θ, and then operating the temporal correlation model _#Λ ( ) on the estimate. Thus the predictor can be realized as:

L = _8/, (θ{SSR, )) = E_ι P_A(e,,,fassκ,)) tø«^] (9)

For the remainder of this section we refer to the predictor obtained from equation (7) as Model based predictor, that obtained from equation (8) as SSR based predictor and that that obtained from (9) as SSR + Model based predictor. It is readily understood that other type of observable parameters, including B, may be used as side-information.

[0036] With reference to Figure 7, this method for predicting the BER of a given data packet may be summarized as follows. Another data packet temporally correlated to the given data packet is received 71 at a receiver. For instance, the another data packet immediately precedes the given data packet.

A bit error rate for the another data packet is estimated at 72 using variable freely observable in the manner described above. Other techniques for estimating the bit error rate of the another data packet are also within the broader aspects of this disclosure. The bit error rate for the given data packet is then determined at 73 using a model that predicts the bit error rate based on the estimated bit error rate of the temporally correlated data packet.

[0037] As done above, we measure the accuracy of our various predictors by employing EKLD as follows:

EKLD[A] =^-\ ∑ ^DKH_*.)] 0°)

expressed by equation (10) we limit the analysis for our predictors to only predicting errors in the corrupted packets. However, our methods can be easily generalized by developing a packet level model for predicting the event of a packet being corrupted. Also note that in equation (10) N_n just merely represents the number of times we see two consecutive corrupted packets in trace

[0038] We are also interested in determining the loss in accuracy (if any) on account employing SSR based predictor or SSR + Model based predictor instead of the Model based predictor. Thus we measure the concentration loss as:

concentration loss = 10 • log_I0 [ ^EKLD»°'*}^{> h})\ (11 )

^Bl0[ EKLD_M [A] J

where the term Model specifically refers to the predictor g_A (ø.) obtained by training on trace Λ , while Me[SSR, SSR+ Model} . It is important to note that the SSR models are obtained by training on all the traces.

[0039] Figure 8A and 8B show the concentration loss 5.5Mbps and 1 1 Mbps respectively. Ideally we would like to limit the concentration loss to less than 3db, at worst to less than 5dB (which represents a loss of concentration by a factor of 3). It can be seen that when we use the SSR based predictor the concentration loss, for 14/20 traces at 11 Mbps and for 12/20 traces at 5.5Mbps, is less than 5dB. These numbers drop to 9 and 6 when we want to limit the concentration loss to approx. 3db or less. Thus even though SSR based predictor can often provide satisfactory prediction, we would desire to have mechanism that provides robust prediction more consistently. Hence, to achieve additional gains a link-specific model can be utilized in conjunction with SSR. It can be observed in Figure 8 that for SSR + Model based predictor the concentration loss is less than 3dB for 15/20 traces at both 5.5Mbps as well 1 1 MBps. Thus the novel mechanism of combining a global side-information based model with a link-specific temporal correlation model, suggested in this work, can indeed lead to significant performance benefits. In fact it can be seen that for 10 traces at 5.5 Mbps and for 8 traces at 11 Mbps the concentration loss is within 0.15dB.

[0040] Measurement based studies are often important pre-cursors to design of efficient communication protocols. The past few years have seen some excellent comprehensive studies of 802.1 1 b networks. However almost all of these studies limit their measurements to packet losses and are void of any measurement/analysis/modeling of bit errors. In this work we investigated the feasibility of utilizing observable variables in CSI and CSP. It was observed that both BT and SSR can be used to improve the accuracy of CSI. The gains provided by BT when used by itself or when used along with SSR are evident but not always significant. Thus there is enough promise in utilizing BT as side- information but for improved performance finer measurements of BT may be essential. As against this the gains provided by SSR in CSI are more significant and consistent. These gains were observed on a variety of traces despite the presence of large amount of BT. Experiments showed that the suggested CSI mechanism can be efficiently combined with the temporal correlations in the BER to facilitate CSP. We suggested a mechanism by which it is feasible to combine a global SSR based CSI mechanism with BER based link-specific temporal correlation model to facilitate CSP. This mechanism was observed to provide almost as good a performance as would be achievable if the BER in a packet is exactly observable.

[0041] The above description is merely exemplary in nature and is not intended to limit the present disclosure, application, or uses.

Claims

CLAIMS What is claimed is:

1. A method for improving error recovery of corrupt data packets in a wireless communication system, comprising: receiving at a receiver a plurality of incoming data packets sent over a wireless medium by an identified transmitter; determining a metric indicative of background traffic observed by the receiver; and estimating a bit error rate for a given data packet based in part on the metric.

2. The method of claim 1 wherein determining a metric further comprises determining a number of data packets received at the receiver and sent by transmitters other than the identified transmitter.

3. The method of claim 1 wherein determining a metric further comprises counting a number of data packets received between two data packets sent by the identified transmitter and dividing by a time over which the number of data packets were received.

4. The method of claim 1 further comprises determining a metric indicative of background traffic at a data link layer of the receiver.

5. The method of claim 1 further comprises determining a signal to silence ratio for the given data packet and estimating the bit error rate for the given data packet based on the signal to silence ratio for the given data packet.

6. The method of claim 1 further comprises passing the bit error rate for the given data packet to the application layer of the receiver; and performing an error recovery operation in relation to the given data packet at the application layer using the bit error rate associated with the given data packet

7. A method for deriving a model for estimating bit error rate of corrupt data packets, comprising: providing a plurality of data packets for deriving a model, where each data packet is labeled with at least one observable parameter and a bit error rate associated with the data packet; grouping data packets having similar observable parameters; defining a cost function for estimating bit error rate based on the observable parameter; and minimizing the cost function for data packets in a given group of data packets, thereby determining an estimated bit error rate for data packets having the observable parameter associated with the given group of data packets.

8. The method of claim 10 further comprises defining the cost function as a Kullback-Leibler divergence measure of bit error rate.

9. The method of claim 10 further comprises grouping the data packets based on a signal to silence ratio associated with the data packets

10. The method of claim 10 further comprises grouping the data packets based on a metric indicative of background traffic observed by a receiver of the data packets.

11. A method for deriving a model for estimating a parameter associated with a data packet, comprising: providing a plurality of data packets for deriving a model, where each data packet is labeled with at least one observable parameter and an unobservable parameter associated with the data packet; grouping data packets having similar observable parameters; defining a cost function for estimating bit error rate based on the observable parameter; and minimizing the cost function for data packets in a given group of data packets, thereby determining an estimate for the unobservable parameter for data packets having the observable parameter associated with the given group of data packets.

12. The method of claim 1 1 further comprises defining the cost function as a Kullback-Leibler divergence measure of bit error rate.

13. The method of claim 1 1 further comprises minimizing the cost function for each group of data packets to derive a set of estimates for the unobservable parameter.

14. A method for improving error recovery of corrupt data packets in a wireless communication system, comprising: receiving a given data packet sent over a wireless medium at a receiver; estimating a bit error rate for another data packet using variables freely observable at a data link layer of the receiver, wherein the another data packet is temporally correlated to the given data packet; and determining a bit error rate for the given data packet using a model that predicts the bit error rate of the given data packet based on the estimated bit error rate of the temporally correlated data packet.

15. The method of claim 14 wherein the another data packet is adjacent to the given data packet.

16. The method of claim 14 wherein the another data packet precedes the given data packet.

17. The method of claim 14 further comprises estimating a bit error rate for the another data packet using a signal to silence ratio for the another data packet.

18. The method of claim 14 further comprises estimating a bit error rate for the another data packet using a metric indicative of background traffic observed by the receiver.

19. The method of claim 14 further comprises determining the metric by counting a number of data packets received between two data packets sent from an expected transmitter.

20. The method of claim 14 further comprises determining a bit error rate for the given data packet using a Markov model.

21. The method of claim 14 further comprises passing the bit error rate for the given data packet to an application layer of the receiver; and performing an error recovery operation in relation to the given data packet at the application layer using the bit error rate associated with the given data packet.

22. The method of claim 14 further comprises communicating the bit error rate for the given data packet to a transmitter of the given data packet and controlling rate of source and channel codes based on the bit error rate of the given data packet.