Search

Finding optimal policies for Partially Observable Markov Decision Processes (POMDPs) is challenging due to their uncountable state spaces when transformed into fully observable Markov Decision Processes (MDPs) using belief states. Traditional methods such as dynamic programming or policy iteration are difficult to apply in this context, necessitating the use of approximation methods on belief states or other techniques. Recently, in (Journal of Machine Learning Research, vol. 23, pp. 1-46, 2022) and (Mathematics of Operations Research, vol. 48, pp. 2066-2093, Nov. 2023), it was shown that sliding finite window based policies are near-optimal for POMDPs with standard Borel valued hidden state spaces, and can be learned via reinforcement learning, with error bounds explicitly dependent on a uniform filter stability term involving total variation in expectation and sample path-wise, respectively. In this paper, we refine these performance bounds and (i) extend them to bounds via uniform filter stability in expected Wasserstein distance leading to an error bound in expectation, and (ii) complementary conditions bounds via uniform filter stability in sample path-wise total variation distance leading to a uniform error bound. We present explicit examples. Our approach thus provides complementary and more refined bounds on the error terms in both total variation and Wasserstein metrics, offering more relaxed and stronger bounds over the approximation error in POMDP solutions on the performance and near optimality of sliding finite window control policies.

The paper focuses on mean-field type multi-agent control problems where the dynamics and cost structures are symmetric and homogeneous, and are affected by the distribution of the agents. A standard solution method for these problems is to consider the infinite population limit as an approximation and use symmetric solutions of the limit problem to achieve near optimality. The control policies, and in particular the dynamics, depend on the population distribution in the finite population setting, or the marginal distribution of the state variable of a representative agent for the infinite population setting. Hence, learning and planning for these control problems generally require estimating the reaction of the system to all possible state distributions of the agents. To overcome this issue, we consider linear function approximation for the control problem and provide several coordinated and independent learning methods. We rigorously establish error upper bounds for the performance of learned solutions. The performance gap stems from (i) the mismatch due to estimating the true model with a linear one, and (ii) using the infinite population solution in the finite population problem as an approximate control. The provided upper bounds quantify the impact of these error sources on the overall performance.

The average cost optimality is known to be a challenging problem for partially observable stochastic control, with few results available beyond the finite state, action, and measurement setup, for which somewhat restrictive conditions are available. In this paper, we present explicit and easily testable conditions for the existence of solutions to the average cost optimality equation where the state space is compact. In particular, we present a new contraction based analysis, which is new to the literature to our knowledge, building on recent regularity results for non-linear filters. Beyond establishing existence, we also present several implications of our analysis that are new to the literature: (i) robustness to incorrect priors (ii) near optimality of policies based on quantized approximations, (iii) near optimality of policies with finite memory, and (iv) convergence in Q-learning. In addition to our main theorem, each of these represents a novel contribution for average cost criteria., Comment: 26 pages

As a primary contribution, we present a convergence theorem for stochastic iterations, and in particular, Q-learning iterates, under a general, possibly non-Markovian, stochastic environment. Our conditions for convergence involve an ergodicity and a positivity criterion. We provide a precise characterization on the limit of the iterates and conditions on the environment and initializations for convergence. As our second contribution, we discuss the implications and applications of this theorem to a variety of stochastic control problems with non-Markovian environments involving (i) quantized approximations of fully observed Markov Decision Processes (MDPs) with continuous spaces (where quantization break down the Markovian structure), (ii) quantized approximations of belief-MDP reduced partially observable MDPS (POMDPs) with weak Feller continuity and a mild version of filter stability (which requires the knowledge of the model by the controller), (iii) finite window approximations of POMDPs under a uniform controlled filter stability (which does not require the knowledge of the model), and (iv) for multi-agent models where convergence of learning dynamics to a new class of equilibria, subjective Q-learning equilibria, will be studied. In addition to the convergence theorem, some implications of the theorem above are new to the literature and others are interpreted as applications of the convergence theorem. Some open problems are noted., Comment: 2 figures

We study mean-field control problems in discrete-time under the infinite horizon average cost optimality criteria. We focus on both the finite population and the infinite population setups. We show the existence of a solution to the average cost optimality equation (ACOE) and the existence of optimal stationary Markov policies for finite population problems under (i) a minorization condition that provides geometric ergodicity on the collective state process of the agents, and (ii) under standard Lipschitz continuity assumptions on the stage-wise cost and transition function of the agents when the Lipschitz constant of the transition function satisfies a certain bound. For the infinite population problem, we establish the existence of a solution to the ACOE, and the existence of optimal policies under the continuity assumptions on the cost and the transition functions. Finally, we relate the finite population and infinite population control problems: (i) we prove that the optimal value of the finite population problem converges to the optimal value of the infinite population problem as the number of agents grows to infinity; (ii) we show that the accumulation points of the finite population optimal solution corresponds to an optimal solution for the infinite population problem, and finally (iii), we show that one can use the solution of the infinite population problem for the finite population problem symmetrically across the agents to achieve near optimal performance when the population is sufficiently large.

For infinite-horizon average-cost criterion problems, there exist relatively few rigorous approximation and reinforcement learning results. In this paper, for such problems, we present several approximation and reinforcement learning results for Markov Decision Processes with standard Borel spaces. Toward this end, (i) we first provide a discretization based approximation method for fully observed Markov Decision Processes (MDPs) with continuous spaces under average cost criteria, and we provide error bounds for the approximations when the dynamics are only weakly continuous (for asymptotic convergence of errors as the grid sizes vanish) or Wasserstein continuous (with a rate in approximation as the grid sizes vanish) under certain ergodicity assumptions. In particular, we relax the total variation condition given in prior work to weak continuity as well as Wasserstein continuity conditions. (ii) We provide synchronous and asynchronous Q-learning algorithms for continuous spaces via quantization (where the quantized state is taken to be the actual state in corresponding Q-learning algorithms presented in the paper), and establish their convergence; for the former we utilize a span semi-norm approach and for the latter we use a direct contraction approach. (iii) We finally show that the convergence is to the optimal Q values of the finite approximate models constructed via quantization, which implies near optimality of the arrived solution. Our Q-learning convergence results and their convergence to near optimality are new for continuous spaces, and the proof method is new even for finite spaces, to our knowledge., Comment: 3 figures

We study a multi-agent mean field type control problem in discrete time where the agents aim to find a socially optimal strategy and where the state and action spaces for the agents are assumed to be continuous. The agents are only weakly coupled through the distribution of their state variables. The problem in its original form can be formulated as a classical Markov decision process (MDP), however, this formulation suffers from several practical difficulties. In this work, we attempt to overcome the curse of dimensionality, coordination complexity between the agents, and the necessity of perfect feedback collection from all the agents (which might be hard to do for large populations.) We provide several approximations: we establish the near optimality of the action and state space discretization of the agents under standard regularity assumptions for the considered formulation by constructing and studying the measure valued MDP counterpart for finite and infinite population settings. It is a well known approach to consider the infinite population problem for mean-field type models, since it provides symmetric policies for the agents which simplifies the coordination between the agents. However, the optimality analysis is harder as the state space of the measure valued infinite population MDP is continuous (even after space discretization of the agents). Therefore, as a final step, we provide further approximations for the infinite population problem by focusing on smaller sized sub-population distributions.

We study a Q learning algorithm for continuous time stochastic control problems. The proposed algorithm uses the sampled state process by discretizing the state and control action spaces under piece-wise constant control processes. We show that the algorithm converges to the optimality equation of a finite Markov decision process (MDP). Using this MDP model, we provide an upper bound for the approximation error for the optimal value function of the continuous time control problem. Furthermore, we present provable upper-bounds for the performance loss of the learned control process compared to the optimal admissible control process of the original problem. The provided error upper-bounds are functions of the time and space discretization parameters, and they reveal the effect of different levels of the approximation: (i) approximation of the continuous time control problem by an MDP, (ii) use of piece-wise constant control processes, (iii) space discretization. Finally, we state a time complexity bound for the proposed algorithm as a function of the time and space discretization parameters., Comment: 3 figures

Reinforcement learning algorithms often require finiteness of state and action spaces in Markov decision processes (MDPs) (also called controlled Markov chains) and various efforts have been made in the literature towards the applicability of such algorithms for continuous state and action spaces. In this paper, we show that under very mild regularity conditions (in particular, involving only weak continuity of the transition kernel of an MDP), Q-learning for standard Borel MDPs via quantization of states and actions (called Quantized Q-Learning) converges to a limit, and furthermore this limit satisfies an optimality equation which leads to near optimality with either explicit performance bounds or which are guaranteed to be asymptotically optimal. Our approach builds on (i) viewing quantization as a measurement kernel and thus a quantized MDP as a partially observed Markov decision process (POMDP), (ii) utilizing near optimality and convergence results of Q-learning for POMDPs, and (iii) finally, near-optimality of finite state model approximations for MDPs with weakly continuous kernels which we show to correspond to the fixed point of the constructed POMDP. Thus, our paper presents a very general convergence and approximation result for the applicability of Q-learning for continuous MDPs.

In this paper, for POMDPs, we provide the convergence of a Q learning algorithm for control policies using a finite history of past observations and control actions, and, consequentially, we establish near optimality of such limit Q functions under explicit filter stability conditions. We present explicit error bounds relating the approximation error to the length of the finite history window. We establish the convergence of such Q-learning iterations under mild ergodicity assumptions on the state process during the exploration phase. We further show that the limit fixed point equation gives an optimal solution for an approximate belief-MDP. We then provide bounds on the performance of the policy obtained using the limit Q values compared to the performance of the optimal policy for the POMDP, where we also present explicit conditions using recent results on filter stability in controlled POMDPs. While there exist many experimental results, (i) the rigorous asymptotic convergence (to an approximate MDP value function) for such finite-memory Q-learning algorithms, and (ii) the near optimality with an explicit rate of convergence (in the memory size) are results that are new to the literature, to our knowledge.

In the theory of Partially Observed Markov Decision Processes (POMDPs), existence of optimal policies have in general been established via converting the original partially observed stochastic control problem to a fully observed one on the belief space, leading to a belief-MDP. However, computing an optimal policy for this fully observed model, and so for the original POMDP, using classical dynamic or linear programming methods is challenging even if the original system has finite state and action spaces, since the state space of the fully observed belief-MDP model is always uncountable. Furthermore, there exist very few rigorous value function approximation and optimal policy approximation results, as regularity conditions needed often require a tedious study involving the spaces of probability measures leading to properties such as Feller continuity. In this paper, we study a planning problem for POMDPs where the system dynamics and measurement channel model are assumed to be known. We construct an approximate belief model by discretizing the belief space using only finite window information variables. We then find optimal policies for the approximate model and we rigorously establish near optimality of the constructed finite window control policies in POMDPs under mild non-linear filter stability conditions and the assumption that the measurement and action sets are finite (and the state space is real vector valued). We also establish a rate of convergence result which relates the finite window memory size and the approximation error bound, where the rate of convergence is exponential under explicit and testable exponential filter stability conditions. While there exist many experimental results and few rigorous asymptotic convergence results, an explicit rate of convergence result is new in the literature, to our knowledge.

We study continuity and robustness properties of infinite-horizon average expected cost problems with respect to (controlled) transition kernels, and applications of these results to the problem of robustness of control policies designed for approximate models applied to actual systems. We show that sufficient conditions presented in the literature for discounted-cost problems are in general not sufficient to ensure robustness for average-cost problems. However, we show that the average optimal cost is continuous in the convergences of controlled transition kernel models where convergence of models entails (i) continuous weak convergence in state and actions, and (ii) continuous setwise convergence in the actions for every fixed state variable, in addition to either uniform ergodicity or some regularity conditions. We establish that the mismatch error due to the application of a control policy designed for an incorrectly estimated model to the true model decreases to zero as the incorrect model approaches the true model under the stated convergence criteria. Our findings significantly relax related studies in the literature which have primarily considered the more restrictive total variation convergence criteria. Applications to robustness to models estimated through empirical data (where almost sure weak convergence criterion typically holds, but stronger criteria do not) are studied and conditions for asymptotic robustness to data-driven learning are established., Comment: Presented at Conference on Decision and Control 2019. arXiv admin note: text overlap with arXiv:1803.06046

We present a new point-based approach for modeling the appearance of real scenes. The approach uses a raw point cloud as the geometric representation of a scene, and augments each point with a learnable neural descriptor that encodes local geometry and appearance. A deep rendering network is learned in parallel with the descriptors, so that new views of the scene can be obtained by passing the rasterizations of a point cloud from new viewpoints through this network. The input rasterizations use the learned descriptors as point pseudo-colors. We show that the proposed approach can be used for modeling complex scenes and obtaining their photorealistic views, while avoiding explicit surface estimation and meshing. In particular, compelling results are obtained for scene scanned using hand-held commodity RGB-D sensors as well as standard RGB cameras even in the presence of objects that are challenging for standard mesh-based modeling.

We present a system for learning full-body neural avatars, i.e. deep networks that produce full-body renderings of a person for varying body pose and camera position. Our system takes the middle path between the classical graphics pipeline and the recent deep learning approaches that generate images of humans using image-to-image translation. In particular, our system estimates an explicit two-dimensional texture map of the model surface. At the same time, it abstains from explicit shape modeling in 3D. Instead, at test time, the system uses a fully-convolutional network to directly map the configuration of body feature points w.r.t. the camera to the 2D texture coordinates of individual pixels in the image frame. We show that such a system is capable of learning to generate realistic renderings while being trained on videos annotated with 3D poses and foreground masks. We also demonstrate that maintaining an explicit texture representation helps our system to achieve better generalization compared to systems that use direct image-to-image translation.

Weak Feller property of controlled and control-free Markov chains lead to many desirable properties. In control-free setups this leads to the existence of invariant probability measures for compact spaces and applicability of numerical approximation methods. For controlled setups, this leads to existence and approximation results for optimal control policies. We know from stochastic control theory that partially observed systems can be converted to fully observed systems by replacing the original state space with a probability measure-valued state space, with the corresponding kernel acting on probability measures known as the non-linear filter (belief) process. Establishing sufficient conditions for the weak Feller property for such processes is a significant problem, studied under various assumptions and setups in the literature. In this paper, we prove the weak Feller property of the non-linear filter process (i) first under weak continuity of the transition probability of controlled Markov chain and total variation continuity of its observation channel, and then, (ii) under total variation continuity of the transition probability of controlled Markov chain. The former result (i) has first appeared in Feinberg et. al. [Math. Oper. Res. 41(2) (2016) 656-681]. Here, we present a concise and easy to follow alternative proof for this existing result. The latter result (ii) establishes weak Feller property of non-linear filter process under conditions, which have not been previously reported in the literature., Comment: Some of the results in this paper are to be presented at the 2019 IEEE Conference on Decision and Control

In stochastic control applications, typically only an ideal model (controlled transition kernel) is assumed and the control design is based on the given model, raising the problem of performance loss due to the mismatch between the assumed model and the actual model. Toward this end, we study continuity properties of discrete-time stochastic control problems with respect to system models (i.e., controlled transition kernels) and robustness of optimal control policies designed for incorrect models applied to the true system. We study both fully observed and partially observed setups under an infinite horizon discounted expected cost criterion. We show that continuity and robustness cannot be established under weak and setwise convergences of transition kernels in general, but that the expected induced cost is robust under total variation. By imposing further assumptions on the measurement models and on the kernel itself (such as continuous convergence), we show that the optimal cost can be made continuous under weak convergence of transition kernels as well. Using these continuity properties, we establish convergence results and error bounds due to mismatch that occurs by the application of a control policy which is designed for an incorrectly estimated system model to a true model, thus establishing positive and negative results on robustness.Compared to the existing literature, we obtain strictly refined robustness results that are applicable even when the incorrect models can be investigated under weak convergence and setwise convergence criteria (with respect to a true model), in addition to the total variation criteria. These entail positive implications on empirical learning in (data-driven) stochastic control since often system models are learned through empirical training data where typically weak convergence criterion applies but stronger convergence criteria do not., Comment: Conference version to appear at the 2018 IEEE CDC with title "Robustness to Incorrect System Models in Stochastic Control and Application to Data-Driven Learning". The paper is to appear in SIAM J. on Control and Optimization

We study the continuity properties of optimal solutions to stochastic control problems with respect to initial probability measures and applications of these to the robustness of optimal control policies applied to systems with incomplete or incorrect priors. It is shown that for single and multi-stage optimal cost problems, continuity and robustness cannot be established under weak convergence or Wasserstein convergence in general, but that the optimal cost is continuous in the priors under the convergence in total variation under mild conditions. By imposing further assumptions on the measurement models, robustness and continuity also hold under weak convergence of priors. We thus obtain robustness results and bounds on the mismatch error that occurs due to the application of a control policy which is designed for an incorrectly estimated prior in terms of a distance measure between the true prior and the incorrect one. Positive and negative practical implications of these results in empirical learning for stochastic control will be presented, where almost surely weak convergence of i.i.d. empirical measures occurs but stronger notions of convergence, such as total variation convergence, in general, do not.