1 Introduction
Within recent years, a number of informationtheoretic approaches have emerged as practical alternatives to traditional machine learning algorithms. Noteworthy examples include the compressionbased approaches of
Frank, Chui, and Witten (2000) and Bratko et al. (2006) to classification, and Cilibrasi and Vitányi (2005) to clustering. What differentiates these techniques from more traditional machine learning approaches is that they rely on the ability to compress the raw input, rather than combining or learning features relevant to the task at hand. Thus this family of techniques has proven most successful in situations where the nature of the data makes it somewhat unwieldy to specify or learn appropriate features. This class of methods can be formally justified by appealing to various notions within algorithmic information theory, such as Kolmogorov complexity (Li and Vitányi, 2008). In this paper we show how a similarly inspired approach can be applied to reinforcement learning, or more specifically, to the tasks of policy evaluation and onpolicy control.Policy evaluation refers to the task of estimating the value function associated with a given policy, for an arbitrary given environment. The performance of wellknown reinforcement learning techniques such as policy iteration (Howard, 1960), approximate dynamic programming (Bertsekas and Tsitsiklis, 1996; Powell, 2011) and actorcritic methods (Sutton and Barto, 1998), for example, all crucially depend on how well policy evaluation can be performed. In this paper we introduce a modelbased approach to policy evaluation, which transforms the task of estimating a value function to that of learning a particular kind of probabilistic state model.
To better put our work into context, it is worth making the distinction between two fundamentally different classes of model based reinforcement learning methods. Simulation based techniques involve learning some kind of forward model of the environment from which future samples can be generated. Given access to such models, planning can be performed directly using search. Noteworthy recent examples include the work of DoshiVelez (2009), Walsh, Goschin, and Littman (2010), Veness et al. (2010), Veness et al. (2011), Asmuth and Littman (2011), Guez, Silver, and Dayan (2012), Hamilton, Fard, and Pineau (2013) and Tziortziotis, Dimitrakakis, and Blekas (2014). Although the aforementioned works demonstrate quite impressive performance on small domains possessing complicated dynamics, scaling these methods to large state or observation spaces has proven challenging. The main difficulty that arises when using learnt forward models is that the modeling errors tend to compound when reasoning over long time horizons (Talvitie, 2014).
In contrast, another family of techniques, referred to in the literature as planning as inference, attempt to sidestep the issue of needing to perform accurate simulations by reducing the planning task to one of probabilistic inference within a generative model of the system. These ideas have been recently explored in both the neuroscience (Botvinick and Toussaint, 2012; Solway and Botvinick, 2012) and machine learning (Attias, 2003; Poupart, Lang, and Toussaint, 2011)
literature. The experimental results to date have been somewhat inconclusive, making it far from clear whether the transformed problem is any easier to solve in practice. Our main contribution in this paper is to show how to set up a particularly tractable form of inference problem by generalizing compressionbased classification to reinforcement learning. The key novelty is to focus the modeling effort on learning the stationary distribution of a particular kind of augmented Markov chain describing the system, from which we can approximate a type of
dual representation (Wang, Bowling, and Schuurmans, 2007; Wang et al., 2008) of the value function. Using this technique, we were able to produce effective controllers on a problem domain orders of magnitude larger than what has previously been addressed with simulation based methods.2 Background
We start with a brief overview of the parts of reinforcement learning and information theory needed to describe our work, before reviewing compressionbased classification.
2.1 Markov Decision Processes
A Markov Decision Process (MDP) is a type of probabilistic model widely used within reinforcement learning
(Sutton and Barto, 1998; Szepesvári, 2010) and control (Bertsekas and Tsitsiklis, 1996). In this work, we limit our attention to finite horizon, time homogenous MDPs whose action and state spaces are finite. Formally, an MDP is a triplet , where is a finite, nonempty set of states, is a finite, nonempty set of actions andis the transition probability kernel that assigns to each stateaction pair
a probability measure over . and are known as the state space and action space respectively. The transition probability kernel gives rise to the state transition kernel , which gives the probability of transitioning from state to state if action is taken in .An agent’s behavior is determined by a policy, that defines, for each state and time , a probability measure over denoted by . A stationary policy is a policy which is independent of time, which we will denote by where appropriate. At each time , the agent communicates an action to the system in state . The system then responds with a statereward pair . Here we will assume that each reward is bounded between and that the system starts in a state
and executes for an infinite number of steps. Thus the execution of the system can be described by a sequence of random variables
.The finite horizon return from time is defined as . The expected horizon return from time , also known as the value function, is denoted by . The return space is the set of all possible returns. The actionvalue function is defined by . An optimal policy, denoted by , is a policy that maximizes the expected return for all ; in our setting, a statedependent deterministic optimal policy always exists.
2.2 Compression and Sequential Prediction
We now review sequential probabilistic prediction in the context of statistical data compression. An alphabet is a set of symbols. A string of data of length is denoted by . The prefix of , , is denoted by or . The empty string is denoted by . The concatenation of two strings and is denoted by .
A coding distribution is a sequence of probability mass functions , which for all satisfy the constraint that for all , with the base case . From here onwards, whenever the meaning is clear from the argument to , the subscript on will be dropped. Under this definition, the conditional probability of a symbol given previous data is defined as provided
, with the familiar chain rules
and now following.A binary source code assigns to each possible data sequence a binary codeword of length . The typical goal when constructing a source code is to minimize the lengths of each codeword while ensuring that the original data sequence is always recoverable from . A fundamental technique known as arithmetic encoding (Witten, Neal, and Cleary, 1987) makes explicit the connection between coding distributions and source codes. Given a coding distribution and a data sequence , arithmetic encoding constructs a code which produces a binary codeword whose length is essentially . We refer the reader to the standard text of Cover and Thomas (1991) for further information.
2.3 Compressionbased classification
Compressionbased classification was introduced by Frank, Chui, and Witten (2000). Given a sequence of labeled i.i.d. training examples , where and are the input and class labels respectively, one can apply Bayes rule to express the probability of a new example
being classified as class
given the training examples by(1) 
The main idea behind compressionbased classification is to model using a coding distribution for the inputs that is trained on the subset of examples from that match class . Well known nonprobabilistic compression methods such as LempelZiv (Ziv and Lempel, 1977) can be used by forming their associated coding distribution , where is the length of the compressed data in bits under compression method . The class probability can be straightforwardly estimated from its empirical frequency or smoothed versions thereof. Thus the overall accuracy of the classifier essentially depends upon how well the inputs can be modeled by the class conditional coding distribution.
Compressionbased classification has both advantages and disadvantages. On one hand, it is straightforward to apply generic compression techniques (including those operating at the bit or character level) to complicated input types such as richly formatted text or DNA strings (Frank, Chui, and Witten, 2000; Bratko et al., 2006). On the other hand, learning a probabilistic model of the input may be significantly more difficult than directly applying standard discriminative classification techniques. Our approach to policy evaluation, which we now describe, raises similar questions.
3 Compression and Control
We now introduce Compress and Control (cnc), our new method for policy evaluation.
3.1 Overview
Policy evaluation is concerned with the estimation of the stateaction value function . Here we assume that the environment is a finite, time homogenous MDP , and that the policy to be evaluated is a stationary Markov policy . To simplify the exposition, we consider the finite horizon case, and assume that all rewards are drawn from a finite set ; later we will discuss how to remove these restrictions.
At a high level, cnc performs policy evaluation by learning a timeindependent stateaction conditional distribution ; the main technical component of our work involves establishing that this timeindependent conditional probability is well defined. Our technique involves constructing a particular kind of augmented Markov chain whose stationary distribution allows for the recovery of . Given this distribution, we can obtain
In the spirit of compressionbased classification, cnc estimates this distribution by using Bayes rule to combine learnt density models of both and . Although it might seem initially strange to learn a model that conditions on the future return, the next section shows how this counterintuitive idea can be made rigorous.
3.2 Transformation
Our goal is to define a transformed chain whose stationary distribution can be marginalized to obtain a distribution over states, actions and the horizon return. We need two lemmas for this purpose. To make these statements precise, we will use some standard terminology from the Markov chain literature; for more detail, we recommend the textbook of Brémaud (1999).
Definition 1.
A Homogenous Markov Chain (HMC) given by over state space is said to be: (AP) aperiodic iff gcd; (PR) positive recurrent iff ; (IR) irreducible iff ; (EA) essentially aperiodic iff gcd. Note also that EA+IR implies AP.
Although the term ergodic is sometimes used to describe particular combinations of these properties (e.g. AP+PR+IR), here we avoid it in favor of being more explicit.
Lemma 1.
Consider a stochastic process over state space that is independent of a sequence of valued random variables in the sense that , and with only depending on and in the sense that and being independent of . Then, if is an (IR/EA/PR) HMC over , then is an (IR/EA/PR) HMC over .
Lemma 1 allows HMC to be augmented to obtain the HMC , where , and denote the action, state and reward at time respectively; see Figure 1 for a graphical depiction of the dependence structure.
The second result allows the HMC to be further augmented to give the snake HMC (Brémaud, 1999). This construction ensures that there is sufficient information within each augmented state to be able to condition on the horizon return.
Lemma 2.
If is an (IR/EA/PR) HMC over state space , then for any , the stochastic process , where , is an (IR/EA/PR) HMC over .
Now if we assume that the HMC defined by and is (IR+EA+PR), Lemmas 1 and 2 imply that there exists a unique stationary distribution over the augmented state space .
Furthermore, if we let and define
, it is clear that there exists a joint distribution
over such that . Hence the probability is well defined, which allows us to express the actionvalue function as(2) 
Finally, by expanding the expectation and applying Bayes rule, Equation 2 can be further rewritten as
(3)  
The cnc approach to policy evaluation involves directly learning the conditional distributions and in Equation 3 from data, and then using these learnt distributions to form a plugin estimate of . Notice that conditions on the return, similar in spirit to prior work on planning as inference (Attias, 2003; Botvinick and Toussaint, 2012; Solway and Botvinick, 2012). The distinguishing property of cnc is that the conditioning is performed with respect to a stationary distribution that has been explicitly constructed to allow for efficient modeling and inference.
3.3 Online Policy Evaluation
We now provide an online algorithm for compressionbased policy evaluation. This will produce, for all times , an estimate of the horizon expected return as a function of the first actionobservationreward triples.
Constructing our estimate involves modeling the probability terms in Equation 3 using two different coding distributions, and respectively; will encode states conditional on returnaction pairs, and will encode returns conditional on actions. Sample states, actions and returns can be generated by directly executing the system ; Provided the HMC is (IR+EA+PR), Lemmas 1 and 2 ensure that the empirical distributions formed from a sufficiently large sample of action/state/return triples will be arbitrarily close to the required conditional probabilities.
Next we describe how the coding distributions are trained. Given a history with , we define the lagged return at any time by . The sequence of the first states occurring in can be mapped to a subsequence denoted by that is defined by keeping only the states . Similarly, a sequence of lagged returns can be mapped to a subsequence formed by keeping only the returns from . Our value estimate at time of taking action in state can now be defined as
(4) 
where
(5) 
approximates the probability of receiving a return of if action is selected in state .
Implementation.
The actionvalue function estimate can be computed efficiently by maintaining buckets, each corresponding to a particular returnaction pair . Each bucket contains an instance of the coding distribution encoding the state sequence . Similarly, buckets containing instances of are created to encode the various return subsequences. This procedure is summarized in Algorithm 1.
3.4 Analysis
We now show that the stateaction estimates defined by Equation 4 are consistent provided that consistent density estimators are used for both and . Also, we will say converges stochastically to 0 with rate if and only if
and will denote this by writing .
Theorem 1.
Given an horizon, finite state space, finite action space, time homogenous MDP and a stationary policy that gives rise to an (IR+EA+PR) HMC, for all , we have that for any state and action that
provided and are consistent estimators of and respectively. Furthermore, if and then .
Next we state consistency results for two types of estimators often used in modelbased reinforcement learning.
Theorem 2.
The frequency estimator when used as either or is a consistent estimator of or respectively for any , , and ; furthermore, the absolute estimation error converges stochastically to 0 with rate .
Note that the above result is essentially tabular, in the sense that each state is treated atomically. The next result applies to a factored application of multialphabet Context Tree Weighting (ctw) (Tjalkens, Shtarkov, and Willems, 1993; Willems, Shtarkov, and Tjalkens, 1995; Veness et al., 2011), which can handle considerably larger state spaces in practice. In the following, we use the notation to refer to the th factor of state .
Theorem 3.
Given a state space that is factored in the sense that , the estimator when used as , is a consistent estimator of for any , , and ; furthermore, the absolute estimation error converges stochastically to 0 at a rate of .
4 Experimental Results
In this section we describe two sets of experiments. The first set is an experimental validation of our theoretical results using a standard policy evaluation benchmark. The second combines cnc with a variety of density estimators and studies the resulting behavior in a large onpolicy control task.
4.1 Policy Evaluation
Our first experiment involves a simplified version of the game of Blackjack (Sutton and Barto, 1998, Section 5.1). In Blackjack, the agent requests cards from the dealer. A game is won when the agent’s card total exceeds the dealer’s own total. We used cnc to estimate the value of the policy that stays if the player’s sum is 20 or 21, and hits in all other cases. A state is represented by the single card held by the dealer, the player’s card total so far, and whether the player holds a usable ace. In total, there are 200 states, two possible actions (hit or stay), and three possible returns (1, 0 and 1). A DirichletMultinomial model with hyperparameters was used for both and .
Figure 2 depicts the estimated MSE and average maximum squared error of over 100,000 episodes; the mean and maximum are taken over all possible stateaction pairs and averaged over 10,000 trials. We also compared cnc to a firstvisit Monte Carlo value estimate (Szepesvári, 2010). The cnc estimate closely tracks the Monte Carlo estimate, even performing slightly better early on due to the smoothing introduced by the use of a Dirichlet prior. As predicted by the analysis in Section 3.4, the MSE decays toward zero.
4.2 Onpolicy Control
Our next set of experiments explored the onpolicy control behavior of cnc under an greedy policy. The purpose of these experiments is to demonstrate the potential of cnc to scale to large control tasks when combined with a variety of different density estimators. Note that Theorem 1 does not apply here: using cnc in this way violates the assumption that is stationary.
Evaluation Platform.
We evaluated cnc using ALE, the Arcade Learning Environment (Bellemare et al., 2013), a reinforcement learning interface to the Atari 2600 video game platform. Observations in ALE consist of frames of 7bit color pixels generated by the Stella Atari 2600 emulator. Although the emulator generates frames at 60Hz, in our experiments we consider time steps that last 4 consecutive frames, following the existing literature (Bellemare, Veness, and Talvitie, 2014; Mnih et al., 2013). We first focused on the game of Pong, which has an action space of and provides a reward of 1 or 1 whenever a point is scored by either the agent or its computer opponent. Episodes end when either player has scored 21 points; as a result, possible scores for one episode range between 21 to 21, with a positive score corresponding to a win for the agent.
Experimental Setup.
We studied four different cnc agents, with each agent corresponding to a different choice of model for ; the Sparse Adapative Dirichlet (sad) estimator (Hutter, 2013) was used for for all agents. Each agent used an greedy policy (Sutton and Barto, 1998) with respect to its current value function estimates. The exploration rate was initialized to 1.0, then decayed linearly to 0.02 over the course of 200,000 time steps. The horizon was set to steps, corresponding to roughly 5 seconds of play. The agents were evaluated over 10 trials, each lasting 2 million steps.
The first model we consider is a factored application of the sad estimator, a count based model designed for large, sparse alphabets. The model divides the screen into regions. The probability of a particular image patch occurring within each region is modeled using a regionspecific sad estimator. The probability assigned to a whole screen is the product of the probabilities assigned to each patch.
The second model is an autoregressive application of logistic regression
(Bishop, 2006), that assigns a probability to each pixel using a shared set of parameters. The product of these perpixel probabilities determines the probability of a screen under this model. The features for each pixel prediction correspond to the pixel’s local context, similar to standard contextbased image compression techniques (Witten, Moffat, and Bell, 1999). The model’s parameters were updated online using Adagrad (Duchi, Hazan, and Singer, 2011). The hyperparameters (including learning rate, choice of context, etc.) were optimized via the random sampling technique of
Bergstra and Bengio (2012).The third model uses the LempelZiv algorithm (Ziv and Lempel, 1977), a dictionarybased compression technique. It works by adapting its internal data structures over time to assign shorter code lengths to more frequently seen substrings of data. For our application, the pixels in each frame were encoded in rowmajor order, by first searching for the longest sequence in the history matching the new data to be compressed, and then encoding a triple that describes the temporal location of the longest match, its length, as well as the next unmatched symbol. This process repeats until no data is left. Recalling Section 2.3, the (implicit) conditional probability of a state under the LempelZiv model can now be obtained by computing
Results.
As depicted in Figure 3 (left), all three models improved their policies over time. By the end of training, two of these models had learnt control policies achieving win rates of approximately 50% in Pong. Over their last 50 episodes of training, the LempelZiv agents averaged 0.09 points per episode (std. error: 1.79) and the factored SAD agents, 3.29 (std. error: 2.49). While the logistic regression agents were less successful (average 17.87, std. error 0.38) we suspect that further training time would significantly improve their performance. Furthermore, all agents ran at realtime or better. These results highlight how cnc can be successfully combined with fundamentally different approaches to density estimation.
We performed one more experiment to illustrate the effects of combining cnc with a more sophisticated density model. We used SkipCTS, a recent Context Tree Weighting derivative, with a context function tailored to the ALE observation space (Bellemare, Veness, and Talvitie, 2014). As shown in Figure 3 (right), cnc combined with SkipCTS learns a nearoptimal policy in Pong. We also compared our method to existing results from the literature (Bellemare et al., 2013; Mnih et al., 2013), although note that the DQN scores, which correspond to a different training regime and do not include Freeway, are included only for illustrative purposes. As shown in Figure 4, cnc can also learn competitive control policies on Freeway and Q*bert.
Interestingly, we found SkipCTS to be insufficiently accurate for effective MCTS planning when used as a forward model, even with enhancements such as double progressive widening (Couëtoux et al., 2011). In particular, our best simulationbased agent did not achieve a score above in Pong, and performed no better than random in Q*bert and Freeway. In comparison, our cnc variants performed significantly better using orders of magnitude less computation. While it would be premature to draw any general conclusions, the cnc approach does appear to be more forgiving of modeling inaccuracies.
5 Discussion and Limitations
The main strength and key limitation of the cnc approach seems to be its reliance on an appropriate choice of density estimator. One could only expect the method to perform well if the learnt models can capture the observational structure specific to high and low return states. Specifying a model can be thus viewed as committing to a particular kind of compressionbased similarity metric over the state space. The attractive part of this approach is that density modeling is a well studied area, which opens up the possibility of bringing in many ideas from machine learning, statistics and information theory to address fundamental questions in reinforcement learning. The downside of course is that density modeling is itself a difficult problem. Further investigation is required to better understand the circumstances under which one would prefer cnc over more traditional modelfree approaches that rely on function approximation to scale to large and complex problems.
So far we have only applied cnc to undiscounted, finite horizon problems with finite action spaces, and more importantly, finite (and rather small) return spaces. This setting is favorable for cnc, since the perstep running time depends on ; in other words, the worst case running time scales no worse than linearly in the length of the horizon. However, even modest changes to the above setting can change the situation drastically. For example, using discounted return can introduce an exponential dependence on the horizon. Thus an important topic for future work is to further develop the cnc approach for large or continuous return spaces. Since the return space is only one dimensional, it would be natural to consider various discretizations of the return space. For example, one could consider a tree based discretization that recursively subdivides the return space into successively smaller halves. A binary tree of depth would produce intervals of even size with an accuracy of . This implies that to achieve an accuracy of at least we would need to set , which should be feasible for many applications. Furthermore, one could attempt to adaptively learn the best discretization (Hutter, 2005a) or approximate Equation 4 using Monte Carlo sampling. These enhancements seem necessary before we could consider applying cnc to the complete suite of ALE games.
6 Closing Remarks
This paper has introduced cnc, an informationtheoretic policy evaluation and onpolicy control technique for reinforcement learning. The most interesting aspect of this approach is the way in which it uses a learnt probabilistic model that conditions on the future return; remarkably, this counterintuitive idea can be justified both in theory and in practice.
While our initial results show promise, a number of open questions clearly remain. For example, so far the cnc value estimates were constructed by using only the Monte Carlo return as the learning signal. However, one of the central themes in Reinforcement Learning is bootstrapping, the idea of constructing value estimates on the basis of other value estimates (Sutton and Barto, 1998). A natural question to explore is whether bootstrapping can be incorporated into the learning signal used by cnc.
For the case of onpolicy control, it would be also interesting to investigate the use of compression techniques or density estimators that can automatically adapt to nonstationary data. A promising line of investigation might be to consider the class of metaalgorithms given by György, Linder, and Lugosi (2012), that can convert any stationary coding distribution into its piecewise stationary extension; efficient algorithms from this class have shown promise for data compression applications, and come with strong theoretical guarantees (Veness et al., 2013). Furthermore, extending the analysis in Section 3.4 to cover the case of onpolicy control or policy iteration (Howard, 1960) would be highly desirable.
Finally, we remark that informationtheoretic perspectives on reinforcement learning have existed for some time; in particular, Hutter (2005b) described a unification of algorithmic information theory and reinforcement learning, leading to the AIXI optimality notion for reinforcement learning agents. Establishing whether any formal connection exists between this body of work and ours is deferred to the future.
Acknowledgments.
We thank Kee Siong Ng, Andras György, Shane Legg, Laurent Orseau and the anonymous reviewers for their helpful feedback on earlier revisions.
References

Asmuth and Littman (2011)
Asmuth, J., and Littman, M. L.
2011.
Learning is planning: near Bayesoptimal reinforcement learning via
MonteCarlo tree search.
In
Uncertainty in Artificial Intelligence (UAI)
, 19–26.  Attias (2003) Attias, H. 2003. Planning by Probabilistic Inference. In Proceedings of the 9th International Workshop on Artificial Intelligence and Statistics.
 Bellemare et al. (2013) Bellemare, M. G.; Naddaf, Y.; Veness, J.; and Bowling, M. 2013. The Arcade Learning Environment: An Evaluation Platform for General Agents. Journal of Artificial Intelligence Research (JAIR) 47:253–279.
 Bellemare, Veness, and Talvitie (2014) Bellemare, M. G.; Veness, J.; and Talvitie, E. 2014. Skip Context Tree Switching. In Proceedings of the ThirtyFirst International Conference on Machine Learning (ICML).
 Bergstra and Bengio (2012) Bergstra, J., and Bengio, Y. 2012. Random search for hyperparameter optimization. Journal of Machine Learning Research (JMLR) 13:281–305.
 Bertsekas and Tsitsiklis (1996) Bertsekas, D. P., and Tsitsiklis, J. N. 1996. NeuroDynamic Programming. Athena Scientific, 1st edition.
 Bishop (2006) Bishop, C. M. 2006. Pattern Recognition and Machine Learning (Information Science and Statistics). Secaucus, NJ, USA: SpringerVerlag New York, Inc.
 Botvinick and Toussaint (2012) Botvinick, M., and Toussaint, M. 2012. Planning as inference. In Trends in Cognitive Sciences 10, 485–588.
 Bratko et al. (2006) Bratko, A.; Cormack, G. V.; R, D.; Filipi, B.; Chan, P.; Lynam, T. R.; and Lynam, T. R. 2006. Spam filtering using statistical data compression models. Journal of Machine Learning Research (JMLR) 7:2673–2698.
 Brémaud (1999) Brémaud, P. 1999. Markov chains : Gibbs fields, Monte Carlo simulation and queues. Texts in applied mathematics. New York, Berlin, Heidelberg: Springer.
 Cilibrasi and Vitányi (2005) Cilibrasi, R., and Vitányi, P. M. B. 2005. Clustering by compression. IEEE Transactions on Information Theory 51:1523–1545.
 Couëtoux et al. (2011) Couëtoux, A.; Hoock, J.B.; Sokolovska, N.; Teytaud, O.; and Bonnard, N. 2011. Continuous upper confidence trees. In Proceedings of the 5th International Conference on Learning and Intelligent Optimization, LION’05, 433–445. SpringerVerlag.
 Cover and Thomas (1991) Cover, T. M., and Thomas, J. A. 1991. Elements of information theory. New York, NY, USA: WileyInterscience.
 DoshiVelez (2009) DoshiVelez, F. 2009. The Infinite Partially Observable Markov Decision Process. In Advances in Neural Information Processing Systems (NIPS) 22.
 Duchi, Hazan, and Singer (2011) Duchi, J.; Hazan, E.; and Singer, Y. 2011. Adaptive subgradient methods for online learning and stochastic optimization. Journal of Machine Learning Research (JMLR) 12:2121–2159.
 Frank, Chui, and Witten (2000) Frank, E.; Chui, C.; and Witten, I. H. 2000. Text categorization using compression models. In Proceedings of Data Compression Conference (DCC), 200–209. IEEE Computer Society Press.
 Guez, Silver, and Dayan (2012) Guez, A.; Silver, D.; and Dayan, P. 2012. Efficient BayesAdaptive Reinforcement Learning using Samplebased Search. In Advances in Neural Information Processing Systems (NIPS) 25.
 György, Linder, and Lugosi (2012) György, A.; Linder, T.; and Lugosi, G. 2012. Efficient Tracking of Large Classes of Experts. IEEE Transactions on Information Theory 58(11):6709–6725.
 Hamilton, Fard, and Pineau (2013) Hamilton, W. L.; Fard, M. M.; and Pineau, J. 2013. Modelling Sparse Dynamical Systems with Compressed Predictive State Representations. In ICML, volume 28 of JMLR Proceedings, 178–186.
 Howard (1960) Howard, R. A. 1960. Dynamic Programming and Markov Processes. MIT Press.

Hutter (2005a)
Hutter, M.
2005a.
Fast nonparametric Bayesian inference on infinite trees.
In Proceedings of 10th International Conference on Artificial Intelligence and Statistics (AISTATS), 144–151.  Hutter (2005b) Hutter, M. 2005b. Universal Artificial Intelligence: Sequential Decisions Based on Algorithmic Probability. Springer.

Hutter (2013)
Hutter, M.
2013.
Sparse adaptive dirichletmultinomiallike processes.
In
Conference on Computational Learning Theory (COLT)
, 432–459.  Li and Vitányi (2008) Li, M., and Vitányi, P. 2008. An Introduction to Kolmogorov Complexity and Its Applications. Springer, third edition.
 Mnih et al. (2013) Mnih, V.; Kavukcuoglu, K.; Silver, D.; Graves, A.; Antonoglou, I.; Wierstra, D.; and Riedmiller, M. 2013. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602.
 Poupart, Lang, and Toussaint (2011) Poupart, P.; Lang, T.; and Toussaint, M. 2011. Escaping Local Optima in POMDP Planning as Inference. In The 10th International Conference on Autonomous Agents and Multiagent Systems  Volume 3, AAMAS ’11, 1263–1264.

Powell (2011)
Powell, W. B.
2011.
Approximate Dynamic Programming: Solving the Curses of Dimensionality
. WileyInterscience, 2nd edition.  Solway and Botvinick (2012) Solway, A., and Botvinick, M. 2012. Goaldirected decision making as probabilistic inference: A computational framework and potential neural correlates. Psycholological Review 119:120–154.
 Sutton and Barto (1998) Sutton, R. S., and Barto, A. G. 1998. Reinforcement learning: An introduction. Cambridge, MA: MIT Press.
 Szepesvári (2010) Szepesvári, C. 2010. Algorithms for Reinforcement Learning. Synthesis Lectures on Artificial Intelligence and Machine Learning. Morgan & Claypool Publishers.
 Talvitie (2014) Talvitie, E. 2014. Model Regularization for Stable Sample Rollouts. In Uncertainty in Artificial Intelligence (UAI).
 Tjalkens, Shtarkov, and Willems (1993) Tjalkens, T. J.; Shtarkov, Y. M.; and Willems, F. M. J. 1993. Context tree weighting: Multialphabet sources. In Proceedings of the 14th Symposium on Information Theory Benelux.
 Tziortziotis, Dimitrakakis, and Blekas (2014) Tziortziotis, N.; Dimitrakakis, C.; and Blekas, K. 2014. Cover Tree Bayesian Reinforcement Learning. Journal of Machine Learning Research (JMLR) 15:2313–2335.
 Veness et al. (2010) Veness, J.; Ng, K. S.; Hutter, M.; and Silver, D. 2010. Reinforcement Learning via AIXI Approximation. In Proceedings of the Conference for the Association for the Advancement of Artificial Intelligence (AAAI).
 Veness et al. (2011) Veness, J.; Ng, K. S.; Hutter, M.; Uther, W.; and Silver, D. 2011. A Monte Carlo AIXI approximation. Journal of Artificial Intelligence Research (JAIR) 40:95–142.
 Veness et al. (2013) Veness, J.; White, M.; Bowling, M.; and Gyorgy, A. 2013. Partition Tree Weighting. In Proceedings of Data Compression Conference (DCC), 321–330.
 Walsh, Goschin, and Littman (2010) Walsh, T. J.; Goschin, S.; and Littman, M. L. 2010. Integrating SampleBased Planning and ModelBased Reinforcement Learning. In Proceedings of the Conference for the Association for the Advancement of Artificial Intelligence (AAAI).
 Wang et al. (2008) Wang, T.; Bowling, M.; Schuurmans, D.; and Lizotte, D. J. 2008. Stable dual dynamic programming. In Advances in Neural Information Processing Systems (NIPS) 20, 1569–1576.
 Wang, Bowling, and Schuurmans (2007) Wang, T.; Bowling, M.; and Schuurmans, D. 2007. Dual representations for dynamic programming and reinforcement learning. In IEEE International Symposium on Approximate Dynamic Programming and Reinforcement Learning, 44–51.
 Willems, Shtarkov, and Tjalkens (1995) Willems, F. M.; Shtarkov, Y. M.; and Tjalkens, T. J. 1995. The Context Tree Weighting Method: Basic Properties. IEEE Transactions on Information Theory 41:653–664.
 Witten, Moffat, and Bell (1999) Witten, I. H.; Moffat, A.; and Bell, T. C. 1999. Managing gigabytes: compressing and indexing documents and images. Morgan Kaufmann.
 Witten, Neal, and Cleary (1987) Witten, I. H.; Neal, R. M.; and Cleary, J. G. 1987. Arithmetic coding for data compression. Communications of the ACM. 30:520–540.
 Ziv and Lempel (1977) Ziv, J., and Lempel, A. 1977. A universal algorithm for sequential data compression. Information Theory, IEEE Transactions on 23(3):337–343.
Comments
There are no comments yet.