Markov games and robust MDPs are closely related models that involve computing a pair of saddle point policies. As part of the long-standing effort to develop efficient algorithms for these models, the Filar-Tolwinski (FT) algorithm has shown considerable promise. As our first contribution, we demonstrate that FT may fail to converge to a saddle point and may loop indefinitely, even in small games. This observation contradicts the proof of FT's optimality in the original paper. As our second contribution, we then propose Residual Conditioned Policy Iteration (RCPI). RCPI builds on FT, but is guaranteed to converge to a saddle point. Our numerical results show that RCPI outperforms other convergent algorithms by several orders of magnitude.

Optimizing risk-averse objectives in discounted MDPs is challenging because most models do not admit direct dynamic programming equations and require complex history-dependent policies. In this paper, we show that the risk-averse total reward criterion, under the Entropic Risk Measure (ERM) and Entropic Value at Risk (EVaR) risk measures, can be optimized by a stationary policy, making it simple to analyze, interpret, and deploy. We propose exponential value iteration, policy iteration, and linear programming to compute optimal policies. Compared with prior work, our results only require the relatively mild condition of transient MDPs and allow for both positive and negative rewards. Our results indicate that the total reward criterion may be preferable to the discounted criterion in a broad range of risk-averse reinforcement learning domains.

In Markov decision processes (MDPs), quantile risk measures such as Value-at-Risk are a standard metric for modeling RL agents’ preferences for certain outcomes. This paper proposes a new Q-learning algorithm for quantile optimization in MDPs with strong convergence and performance guarantees. The algorithm leverages a new, simple dynamic program (DP) decomposition for quantile MDPs. Compared with prior work, our DP decomposition requires neither known transition probabilities nor solving complex saddle point equations and serves as a suitable foundation for other model-free RL algorithms. Our numerical results in tabular domains show that our Q-learning algorithm converges to its DP variant and outperforms earlier algorithms.

Robust Markov Decision Processes (MDPs) offer a promising framework for computing reliable policies under model uncertainty. While policy gradient methods have gained increasing popularity in robust discounted MDPs, their application to the average-reward criterion remains largely unexplored. This paper proposes a Robust Projected Policy Gradient (RP2G), the first generic policy gradient method for robust average-reward MDPs (RAMDPs) that is applicable beyond the typical rectangularity assumption on transition ambiguity. In contrast to existing robust policy gradient algorithms, RP2G incorporates an adaptive decreasing tolerance mechanism for efficient policy updates at each iteration. We also present a comprehensive convergence analysis of RP2G for solving ergodic tabular RAMDPs. Furthermore, we establish the first study of the inner worst-case transition evaluation problem in RAMDPs, proposing two gradient-based algorithms tailored for rectangular and general ambiguity sets, each with provable convergence guarantees. Numerical experiments confirm the global convergence of our new algorithm and demonstrate its superior performance.

Risk-averse total-reward Markov Decision Processes (MDPs) offer a promising framework for modeling and solving undiscounted infinite-horizon objectives. Existing model-based algorithms for risk measures like the entropic risk measure (ERM) and entropic value-at-risk (EVaR) are effective in small problems, but require full access to transition probabilities. We propose a Q-learning algorithm to compute the optimal stationary policy for total-reward ERM and EVaR objectives with strong convergence and performance guarantees. The algorithm and its optimality are made possible by ERM's dynamic consistency and elicitability. Our numerical results on tabular domains demonstrate quick and reliable convergence of the proposed Q-learning algorithm to the optimal risk-averse value function.

We study how to make decisions that minimize Bayesian regret in offline linear bandits. Prior work suggests that one must take actions with maximum lower confidence bound (LCB) on their reward. We argue that reliance on LCB is inherently flawed in this setting and propose a new algorithm that directly minimizes upper-bounds on the Bayesian regret using efficient conic optimization solvers. Our bounds build heavily on new connections to monetary risk measures. Proving a matching lower-bound, we show that our upper-bounds are tight, and by minimizing them we are guaranteed to outperform the LCB approach. Our numerical results on synthetic domains confirm that our approach is superior to maximizing LCB.

Risk-averse Markov Decision Processes (MDPs) have optimal policies that achieve high returns with low variability, but these MDPs are often difficult to solve. Only a few practical risk-averse objectives admit a dynamic programming (DP) formulation, which is the mainstay of most MDP and RL algorithms. We derive a new DP formulation for discounted risk-averse MDPs with Entropic Risk Measure (ERM) and Entropic Value at Risk (EVaR) objectives. Our DP formulation for ERM, which is possible because of our novel definition of value function with time-dependent risk levels, can approximate optimal policies in a time that is polynomial in the approximation error. We then use the ERM algorithm to optimize the EVaR objective in polynomial time using an optimized discretization scheme. Our numerical results show the viability of our formulations and algorithms in discounted MDPs.

Robust Markov decision processes (RMDPs) provide a promising framework for computing reliable policies in the face of model errors. Many successful reinforcement learning algorithms build on variations of policy-gradient methods, but adapting these methods to RMDPs has been challenging. As a result, the applicability of RMDPs to large, practical domains remains limited. This paper proposes a new Double-Loop Robust Policy Gradient (DRPG), the first generic policy gradient method for RMDPs. In contrast with prior robust policy gradient algorithms, DRPG monotonically reduces approximation errors to guarantee convergence to a globally optimal policy in tabular RMDPs. We introduce a novel parametric transition kernel and solve the inner loop robust policy via a gradient-based method. Finally, our numerical results demonstrate the utility of our new algorithm and confirm its global convergence properties.

In reinforcement learning, robust policies for high-stakes decision-making problems with limited data are usually computed by optimizing the percentile criterion. The percentile criterion is optimized by constructing an uncertainty set that contains the true model with high probability and optimizing the policy for the worst model in the set. Since the percentile criterion is non-convex, constructing these sets itself is challenging. Existing works use Bayesian credible regions as uncertainty sets, but they are often unnecessarily large and result in learning overly conservative policies. To overcome these shortcomings, we propose a novel Value-at-Risk based dynamic programming algorithm to optimize the percentile criterion without explicitly constructing any uncertainty sets. Our theoretical and empirical results show that our algorithm implicitly constructs much smaller uncertainty sets and learns less-conservative robust policies.

We introduce the Blackwell discount factor for Markov Decision Processes (MDPs). Classical objectives for MDPs include discounted, average, and Blackwell optimality. Many existing approaches to computing average-optimal policies solve for discount-optimal policies with a discount factor close to $1$, but they only work under strong or hard-to-verify assumptions on the MDP structure such as unichain or ergodicity. We are the first to highlight the shortcomings of the classical definition of Blackwell optimality, which does not lead to simple algorithms for computing Blackwell-optimal policies and overlooks the pathological behaviors of optimal policies as regards the discount factors. To resolve this issue, in this paper, we show that when the discount factor is larger than the Blackwell discount factor $\gamma_{\sf bw}$, all discount-optimal policies become Blackwell- and average-optimal, and we derive a general upper bound on $\gamma_{\sf bw}$. Our upper bound on $\gamma_{\sf bw}$, parametrized by the bit-size of the rewards and transition probabilities of the MDP instance, provides the first reduction from average and Blackwell optimality to discounted optimality, without any assumptions, along with new polynomial-time algorithms. Our work brings new ideas from polynomials and algebraic numbers to the analysis of MDPs. Our results also apply to robust MDPs, enabling the first algorithms to compute robust Blackwell-optimal policies.

Optimizing static risk-averse objectives in Markov decision processes is difficult because they do not admit standard dynamic programming equations common in Reinforcement Learning (RL) algorithms. Dynamic programming decompositions that augment the state space with discrete risk levels have recently gained popularity in the RL community. Prior work has shown that these decompositions are optimal when the risk level is discretized sufficiently. However, we show that these popular decompositions for Conditional-Value-at-Risk (CVaR) and Entropic-Value-at-Risk (EVaR) are inherently suboptimal regardless of the discretization level. In particular, we show that a saddle point property assumed to hold in prior literature may be violated. However, a decomposition does hold for Value-at-Risk and our proof demonstrates how this risk measure differs from CVaR and EVaR. Our findings are significant because risk-averse algorithms are used in high-stake environments, making their correctness much more critical.

Multi-model Markov decision process (MMDP) is a promising framework for computing policies that are robust to parameter uncertainty in MDPs. MMDPs aim to find a policy that maximizes the expected return over a distribution of MDP models. Because MMDPs are NP-hard to solve, most methods resort to approximations. In this paper, we derive the policy gradient of MMDPs and propose CADP, which combines a coordinate ascent method and a dynamic programming algorithm for solving MMDPs. The main innovation of CADP compared with earlier algorithms is to take the coordinate ascent perspective to adjust model weights iteratively to guarantee monotone policy improvements to a local maximum. A theoretical analysis of CADP proves that it never performs worse than previous dynamic programming algorithms like WSU. Our numerical results indicate that CADP substantially outperforms existing methods on several benchmark problems.

In recent years, robust Markov decision processes (MDPs) have emerged as a prominent modeling framework for dynamic decision problems affected by uncertainty. In contrast to classical MDPs, which only account for stochasticity by modeling the dynamics through a stochastic process with a known transition kernel, robust MDPs additionally account for ambiguity by optimizing in view of the most adverse transition kernel from a prescribed ambiguity set. In this paper, we develop a novel solution framework for robust MDPs with $s$-rectangular ambiguity sets that decomposes the problem into a sequence of robust Bellman updates and simplex projections. Exploiting the rich structure present in the simplex projections corresponding to $\phi$-divergence ambiguity sets, we show that the associated $s$-rectangular robust MDPs can be solved substantially faster than with state-of-the-art commercial solvers as well as a recent first-order solution scheme, thus rendering them attractive alternatives to classical MDPs in practical applications.

Off-policy Evaluation (OPE) methods are a crucial tool for evaluating policies in high-stakes domains such as healthcare, where exploration is often infeasible, unethical, or expensive. However, the extent to which such methods can be trusted under adversarial threats to data quality is largely unexplored. In this work, we make the first attempt at investigating the sensitivity of OPE methods to marginal adversarial perturbations to the data. We design a generic data poisoning attack framework leveraging influence functions from robust statistics to carefully construct perturbations that maximize error in the policy value estimates. We carry out extensive experimentation with multiple healthcare and control datasets. Our results demonstrate that many existing OPE methods are highly prone to generating value estimates with large errors when subject to data poisoning attacks, even for small adversarial perturbations. These findings question the reliability of policy values derived using OPE methods and motivate the need for developing OPE methods that are statistically robust to train-time data poisoning attacks.

We address the problem of computing reliable policies in reinforcement learning problems with limited data. In particular, we compute policies that achieve good returns with high confidence when deployed. This objective, known as the percentile criterion, can be optimized using Robust MDPs (RMDPs). RMDPs generalize MDPs to allow for uncertain transition probabilities chosen adversarially from given ambiguity sets. We show that the RMDP solution’s sub-optimality depends on the spans of the ambiguity sets along the value function. We then propose new algorithms that minimize the span of ambiguity sets defined by weighted L1 and L-infinity norms. Our primary focus is on Bayesian guarantees, but we also describe how our methods apply to frequentist guarantees and derive new concentration inequalities for weighted L1 and L-infinity norms. Experimental results indicate that our optimized ambiguity sets improve significantly on prior construction methods.

The difficulty in specifying rewards for many real-world problems has led to an increased focus on learning rewards from human feedback, such as demonstrations. However, there are often many different reward functions that explain the human feedback, leaving agents with uncertainty over what the true reward function is. While most policy optimization approaches handle this uncertainty by optimizing for expected performance, many applications demand risk-averse behavior. We derive a novel policy gradient-style robust optimization approach, PG-BROIL, that optimizes a soft-robust objective that balances expected performance and risk. To the best of our knowledge, PG-BROIL is the first policy optimization algorithm robust to a distribution of reward hypotheses which can scale to continuous MDPs. Results suggest that PG-BROIL can produce a family of behaviors ranging from risk-neutral to risk-averse and outperforms state-of-the-art imitation learning algorithms when learning from ambiguous demonstrations by hedging against uncertainty, rather than seeking to uniquely identify the demonstrator’s reward function.

Imitation learning (IL) frameworks in robotics typically assume that a domain expert's demonstration always contains a correct way of doing the task. Despite its theoretical convenience, this assumption has limited practical values for an IL-powered robot in real world. There are many reasons for an expert in the real world to provide demonstrations that may contain incorrect or potentially unsafe way of doing a task. In order for IL-powered robots to work in the real world, IL frameworks need to detect such adversarial demonstrations and not learn from them. This paper proposes an IL framework that can autonomously detect and remove adversarial demonstrations, if they exist in the demonstration set, as it directly learns a task policy from the expert. The proposed framework that we term Robust Maximum Entropy behavior cloning (R-MaxEnt) learns a stochastic model that maps states to actions. In doing so, R-MaxEnt solves a minmax problem that leverages the entropy of the model to assign weights to different demonstrations while assigning poor weights to adversarial samples. Our empirical results show that R-MaxEnt outperforms the existing IL approaches in both real and simulated robotics tasks.

Robust Markov decision processes (RMDPs) are a useful building block of robust reinforcement learning algorithms but can be hard to solve. This paper proposes a fast, exact algorithm for computing the Bellman operator for S-rectangular robust Markov decision processes with $L_\infty$-constrained rectangular ambiguity sets. The algorithm combines a novel homotopy continuation method with a bisection method to solve S-rectangular ambiguity in quasi-linear time in the number of states and actions. The algorithm improves on the cubic time required by leading general linear programming methods. Our experimental results confirm the practical viability of our method and show that it outperforms a leading commercial optimization package by several orders of magnitude.

Robust Markov decision processes (MDPs) compute reliable solutions for dynamic decision problems with partially-known transition probabilities. Unfortunately, accounting for uncertainty in the transition probabilities significantly increases the computational complexity of solving robust MDPs, which limits their scalability. This paper describes new, efficient algorithms for solving the common class of robust MDPs with s- and sa-rectangular ambiguity sets defined by weighted L1 norms. We propose partial policy iteration, a new, efficient, flexible, and general policy iteration scheme for robust MDPs. We also propose fast methods for computing the robust Bellman operator in quasi-linear time, nearly matching the ordinary Bellman operator's linear complexity. Our experimental results indicate that the proposed methods are many orders of magnitude faster than the state-of-the-art approach, which uses linear programming solvers combined with a robust value iteration.

Suboptimal heuristic search algorithms can benefit from reasoning about heuristic error, especially in a real-time setting where there is not enough time to search all the way to a goal. However, current reasoning methods implicitly or explicitly incorporate assumptions about the cost-to-go function. We consider a recent real-time search algorithm, called Nancy, that manipulates explicit beliefs about the cost-to-go. The original presentation of Nancy assumed that these beliefs are Gaussian, with parameters following a certain form. In this paper, we explore how to replace these assumptions with actual data. We develop a data-driven variant of Nancy, DDNancy, that bases its beliefs on heuristic performance statistics from the same domain. We extend Nancy and DDNancy with the notion of persistence and prove their completeness. Experimental results show that DDNancy can perform well in domains in which the original assumption-based Nancy performs poorly.

One of the main challenges in imitation learning is determining what action an agent should take when outside the state distribution of the demonstrations. Inverse reinforcement learning (IRL) can enable generalization to new states by learning a parameterized reward function, but these approaches still face uncertainty over the true reward function and corresponding optimal policy. Existing safe imitation learning approaches based on IRL deal with this uncertainty using a maxmin framework that optimizes a policy under the assumption of an adversarial reward function, whereas risk-neutral IRL approaches either optimize a policy for the mean or MAP reward function. While completely ignoring risk can lead to overly aggressive and unsafe policies, optimizing in a fully adversarial sense is also problematic as it can lead to overly conservative policies that perform poorly in practice. To provide a bridge between these two extremes, we propose Bayesian Robust Optimization for Imitation Learning (BROIL). BROIL leverages Bayesian reward function inference and a user specific risk tolerance to efficiently optimize a robust policy that balances expected return and conditional value at risk. Our empirical results show that BROIL provides a natural way to interpolate between return-maximizing and risk-minimizing behaviors and outperforms existing risk-sensitive and risk-neutral inverse reinforcement learning algorithms.

This paper presents a framework to learn the reward function underlying high-level sequential tasks from demonstrations. The purpose of reward learning, in the context of learning from demonstration (LfD), is to generate policies that mimic the demonstrator's policies, thereby enabling imitation learning. We focus on a human-robot interaction (HRI) domain where the goal is to learn and model structured interactions between a human and a robot. Such interactions can be modeled as a partially observable Markov decision process (POMDP) where the partial observability is caused by uncertainties associated with the ways humans respond to different stimuli. The key challenge in finding a good policy in such a POMDP is determining the reward function that was observed by the demonstrator. Existing inverse reinforcement learning (IRL) methods for POMDPs are computationally very expensive and the problem is not well understood. In comparison, IRL algorithms for Markov decision process (MDP) are well defined and computationally efficient. We propose an approach of reward function learning for high-level sequential tasks from human demonstrations where the core idea is to reduce the underlying POMDP to an MDP and apply any efficient MDPIRL algorithm. Our extensive experiments suggest that the reward function learned this way generates POMDP policies that mimic the policies of the demonstrator well.

Robust MDPs (RMDPs) can be used to compute policies with provable worst-case guarantees in reinforcement learning. The quality and robustness of an RMDP solution are determined by the ambiguity set---the set of plausible transition probabilities---which is usually constructed as a multi-dimensional confidence region. Existing methods construct ambiguity sets as confidence regions using concentration inequalities which leads to overly conservative solutions. This paper proposes a new paradigm that can achieve better solutions with the same robustness guarantees without using confidence regions as ambiguity sets. To incorporate prior knowledge, our algorithms optimize the size and position of ambiguity sets using Bayesian inference. Our theoretical analysis shows the safety of the proposed method, and the empirical results demonstrate its practical promise.

We describe two efficient, and exact, algorithms for computing Bellman updates in robust Markov decision processes (MDPs). The first algorithm uses a homotopy continuation method to compute updates for L1-constrained s,a-rectangular ambiguity sets. It runs in quasi-linear time for plain L1-norms and also generalizes to weighted L1-norms. The second algorithm uses bisection to compute updates for robust MDPs with s-rectangular ambiguity sets. This algorithm, when combined with the homotopy method, also has a quasi-linear runtime. Unlike previous methods, our algorithms compute the primal solution in addition to the optimal objective value, which makes them useful in policy iteration methods. Our experimental results indicate that the proposed methods are over 1,000 times faster than Gurobi, a state-of-the-art commercial optimization package, for small instances, and the performance gap grows considerably with problem size.

What policy should be employed in a Markov decision process with uncertain parameters? Robust optimization answer to this question is to use rectangular uncertainty sets, which independently reflect available knowledge about each state, and then obtains a decision policy that maximizes expected reward for the worst-case decision process parameters from these uncertainty sets. While this rectangularity is convenient computationally and leads to tractable solutions, it often produces policies that are too conservative in practice, and does not facilitate knowledge transfer between portions of the state space or across related decision processes. In this work, we propose non-rectangular uncertainty sets that bound marginal moments of state-action features defined over entire trajectories through a decision process. This enables generalization to different portions of the state space while retaining appropriate uncertainty of the decision process. We develop algorithms for solving the resulting robust decision problems, which reduce to finding an optimal policy for a mixture of decision processes, and demonstrate the benefits of our approach experimentally.

An important problem in sequential decision-making under uncertainty is to use limited data to compute a safe policy, i.e., a policy that is guaranteed to perform at least as well as a given baseline strategy. In this paper, we develop and analyze a new model-based approach to compute a safe policy when we have access to an inaccurate dynamics model of the system with known accuracy guarantees. Our proposed robust method uses this (inaccurate) model to directly minimize the (negative) regret w.r.t. the baseline policy. Contrary to the existing approaches, minimizing the regret allows one to improve the baseline policy in states with accurate dynamics and seamlessly fall back to the baseline policy, otherwise. We show that our formulation is NP-hard and propose an approximate algorithm. Our empirical results on several domains show that even this relatively simple approximate algorithm can significantly outperform standard approaches.

Sustainable resource management in many domains presents large continuous stochastic optimization problems, which can often be modeled as Markov decision processes (MDPs). To solve such large MDPs, we identify and leverage linearity in state and action sets that is common in resource management. In particular, we introduce linear dynamic programs (LDPs) that generalize resource management problems and partially observable MDPs (POMDPs). We show that the LDP framework makes it possible to adapt point-based methods--the state of the art in solving POMDPs--to solving LDPs. The experimental results demonstrate the efficiency of this approach in managing the water level of a river reservoir. Finally, we discuss the relationship with dual dynamic programming, a method used to optimize hydroelectric systems.

Approximate Linear Programming (ALP) is a reinforcement learning technique with nice theoretical properties, but it often performs poorly in practice. We identify some reasons for the poor quality of ALP solutions in problems where the approximation induces virtual loops. We then introduce two methods for improving solution quality. One method rolls out selected constraints of the ALP, guided by the dual information. The second method is a relaxation of the ALP, based on external penalty methods. The latter method is applicable in domains in which rolling out constraints is impractical. Both approaches show promising empirical results for simple benchmark problems as well as for a realistic blood inventory management problem.

Existing value function approximation methods have been successfully used in many applications, but they often lack useful a priori error bounds. We propose approximate bilinear programming, a new formulation of value function approximation that provides strong a priori guarantees. In particular, it provably finds an approximate value function that minimizes the Bellman residual. Solving a bilinear program optimally is NP hard, but this is unavoidable because the Bellman-residual minimization itself is NP hard. We, therefore, employ and analyze a common approximate algorithm for bilinear programs. The analysis shows that this algorithm offers a convergent generalization of approximate policy iteration. Finally, we demonstrate that the proposed approach can consistently minimize the Bellman residual on a simple benchmark problem.

Multiagent planning and coordination problems are common and known to be computationally hard. We show that a wide range of two-agent problems can be formulated as bilinear programs. We present a successive approximation algorithm that significantly outperforms the coverage set algorithm, which is the state-of-the-art method for this class of multiagent problems. Because the algorithm is formulated for bilinear programs, it is more general and simpler to implement. The new algorithm can be terminated at any time and-unlike the coverage set algorithm-it facilitates the derivation of a useful online performance bound. It is also much more efficient, on average reducing the computation time of the optimal solution by about four orders of magnitude. Finally, we introduce an automatic dimensionality reduction method that improves the effectiveness of the algorithm, extending its applicability to new domains and providing a new way to analyze a subclass of bilinear programs.

Most algorithms for solving Markov decision processes rely on a discount factor, which ensures their convergence. In fact, it is often used in problems with is no intrinsic motivation. In this paper, we show that when used in approximate dynamic programming, an artificially low discount factor may significantly improve the performance on some problems, such as Tetris. We propose two explanations for this phenomenon. Our first justification follows directly from the standard approximation error bounds: using a lower discount factor may decrease the approximation error bounds. However, we also show that these bounds are loose, a thus their decrease does not entirely justify a better practical performance. We thus propose another justification: when the rewards are received only sporadically (as it is the case in Tetris), we can derive tighter bounds, which support a significant performance increase with a decrease in the discount factor.