RL in Latent MDPs is Tractable: Online Guarantees via Off-Policy Evaluation

In many real-world decision problems there is partially observed, hidden or latent information that remains fixed throughout an interaction. Such decision problems can be modeled as Latent Markov Decision Processes (LMDPs), where a latent variable is selected at the beginning of an interaction and is not disclosed to the agent. In the last decade, there has been significant progress in solving LMDPs under different structural assumptions. However, for general LMDPs, there is no known learning algorithm that provably matches the existing lower bound~\cite{kwon2021rl}. We introduce the first sample-efficient algorithm for LMDPs without any additional structural assumptions. Our result builds off a new perspective on the role of off-policy evaluation guarantees and coverage coefficients in LMDPs, a perspective, that has been overlooked in the context of exploration in partially observed environments. Specifically, we establish a novel off-policy evaluation lemma and introduce a new coverage coefficient for LMDPs. Then, we show how these can be used to derive near-optimal guarantees of an optimistic exploration algorithm. These results, we believe, can be valuable for a wide range of interactive learning problems beyond LMDPs, and especially, for partially observed environments.

On Bits and Bandits: Quantifying the Regret-Information Trade-off

In interactive decision-making tasks, information can be acquired by direct interactions, through receiving indirect feedback, and from external knowledgeable sources. We examine the trade-off between the information an agent accumulates and the regret it suffers. We show that information from external sources, measured in bits, can be traded off for regret, measured in reward. We invoke information-theoretic methods for obtaining regret lower bounds, that also allow us to easily re-derive several known lower bounds. We then generalize a variety of interactive decision-making tasks with external information to a new setting. Using this setting, we introduce the first Bayesian regret lower bounds that depend on the information an agent accumulates. These lower bounds also prove the near-optimality of Thompson sampling for Bayesian problems. Finally, we demonstrate the utility of these bounds in improving the performance of a question-answering task with large language models, allowing us to obtain valuable insights.

Tree Search-Based Policy Optimization under Stochastic Execution Delay

The standard formulation of Markov decision processes (MDPs) assumes that the agent's decisions are executed immediately. However, in numerous realistic applications such as robotics or healthcare, actions are performed with a delay whose value can even be stochastic. In this work, we introduce stochastic delayed execution MDPs, a new formalism addressing random delays without resorting to state augmentation. We show that given observed delay values, it is sufficient to perform a policy search in the class of Markov policies in order to reach optimal performance, thus extending the deterministic fixed delay case. Armed with this insight, we devise DEZ, a model-based algorithm that optimizes over the class of Markov policies. DEZ leverages Monte-Carlo tree search similar to its non-delayed variant EfficientZero to accurately infer future states from the action queue. Thus, it handles delayed execution while preserving the sample efficiency of EfficientZero. Through a series of experiments on the Atari suite, we demonstrate that although the previous baseline outperforms the naive method in scenarios with constant delay, it underperforms in the face of stochastic delays. In contrast, our approach significantly outperforms the baselines, for both constant and stochastic delays. The code is available at http://github.com/davidva1/Delayed-EZ .

On the Global Convergence of Policy Gradient in Average Reward Markov Decision Processes

We present the first finite time global convergence analysis of policy gradient in the context of infinite horizon average reward Markov decision processes (MDPs). Specifically, we focus on ergodic tabular MDPs with finite state and action spaces. Our analysis shows that the policy gradient iterates converge to the optimal policy at a sublinear rate of $O\left({\frac{1}{T}}\right),$ which translates to $O\left({\log(T)}\right)$ regret, where $T$ represents the number of iterations. Prior work on performance bounds for discounted reward MDPs cannot be extended to average reward MDPs because the bounds grow proportional to the fifth power of the effective horizon. Thus, our primary contribution is in proving that the policy gradient algorithm converges for average-reward MDPs and in obtaining finite-time performance guarantees. In contrast to the existing discounted reward performance bounds, our performance bounds have an explicit dependence on constants that capture the complexity of the underlying MDP. Motivated by this observation, we reexamine and improve the existing performance bounds for discounted reward MDPs. We also present simulations to empirically evaluate the performance of average reward policy gradient algorithm.

Conservative DDPG -- Pessimistic RL without Ensemble

DDPG is hindered by the overestimation bias problem, wherein its $Q$-estimates tend to overstate the actual $Q$-values. Traditional solutions to this bias involve ensemble-based methods, which require significant computational resources, or complex log-policy-based approaches, which are difficult to understand and implement. In contrast, we propose a straightforward solution using a $Q$-target and incorporating a behavioral cloning (BC) loss penalty. This solution, acting as an uncertainty measure, can be easily implemented with minimal code and without the need for an ensemble. Our empirical findings strongly support the superiority of Conservative DDPG over DDPG across various MuJoCo and Bullet tasks. We consistently observe better performance in all evaluated tasks and even competitive or superior performance compared to TD3 and TD7, all achieved with significantly reduced computational requirements.

Exploration-Driven Policy Optimization in RLHF: Theoretical Insights on Efficient Data Utilization

Reinforcement Learning from Human Feedback (RLHF) has achieved impressive empirical successes while relying on a small amount of human feedback. However, there is limited theoretical justification for this phenomenon. Additionally, most recent studies focus on value-based algorithms despite the recent empirical successes of policy-based algorithms. In this work, we consider an RLHF algorithm based on policy optimization (PO-RLHF). The algorithm is based on the popular Policy Cover-Policy Gradient (PC-PG) algorithm, which assumes knowledge of the reward function. In PO-RLHF, knowledge of the reward function is not assumed and the algorithm relies on trajectory-based comparison feedback to infer the reward function. We provide performance bounds for PO-RLHF with low query complexity, which provides insight into why a small amount of human feedback may be sufficient to get good performance with RLHF. A key novelty is our trajectory-level elliptical potential analysis technique used to infer reward function parameters when comparison queries rather than reward observations are used. We provide and analyze algorithms in two settings: linear and neural function approximation, PG-RLHF and NN-PG-RLHF, respectively.

Improving Token-Based World Models with Parallel Observation Prediction

Motivated by the success of Transformers when applied to sequences of discrete symbols, token-based world models (TBWMs) were recently proposed as sample-efficient methods. In TBWMs, the world model consumes agent experience as a language-like sequence of tokens, where each observation constitutes a sub-sequence. However, during imagination, the sequential token-by-token generation of next observations results in a severe bottleneck, leading to long training times, poor GPU utilization, and limited representations. To resolve this bottleneck, we devise a novel Parallel Observation Prediction (POP) mechanism. POP augments a Retentive Network (RetNet) with a novel forward mode tailored to our reinforcement learning setting. We incorporate POP in a novel TBWM agent named REM (Retentive Environment Model), showcasing a 15.4x faster imagination compared to prior TBWMs. REM attains superhuman performance on 12 out of 26 games of the Atari 100K benchmark, while training in less than 12 hours. Our code is available at \url{https://github.com/leor-c/REM}.

MinMaxMin $Q$-learning

MinMaxMin $Q$-learning is a novel optimistic Actor-Critic algorithm that addresses the problem of overestimation bias ($Q$-estimations are overestimating the real $Q$-values) inherent in conservative RL algorithms. Its core formula relies on the disagreement among $Q$-networks in the form of the min-batch MaxMin $Q$-networks distance which is added to the $Q$-target and used as the priority experience replay sampling-rule. We implement MinMaxMin on top of TD3 and TD7, subjecting it to rigorous testing against state-of-the-art continuous-space algorithms-DDPG, TD3, and TD7-across popular MuJoCo and Bullet environments. The results show a consistent performance improvement of MinMaxMin over DDPG, TD3, and TD7 across all tested tasks.

SQT -- std $Q$-target

Std $Q$-target is a conservative, actor-critic, ensemble, $Q$-learning-based algorithm, which is based on a single key $Q$-formula: $Q$-networks standard deviation, which is an "uncertainty penalty", and, serves as a minimalistic solution to the problem of overestimation bias. We implement SQT on top of TD3/TD7 code and test it against the state-of-the-art (SOTA) actor-critic algorithms, DDPG, TD3 and TD7 on seven popular MuJoCo and Bullet tasks. Our results demonstrate SQT's $Q$-target formula superiority over TD3's $Q$-target formula as a conservative solution to overestimation bias in RL, while SQT shows a clear performance advantage on a wide margin over DDPG, TD3, and TD7 on all tasks.

Prospective Side Information for Latent MDPs

In many interactive decision-making settings, there is latent and unobserved information that remains fixed. Consider, for example, a dialogue system, where complete information about a user, such as the user's preferences, is not given. In such an environment, the latent information remains fixed throughout each episode, since the identity of the user does not change during an interaction. This type of environment can be modeled as a Latent Markov Decision Process (LMDP), a special instance of Partially Observed Markov Decision Processes (POMDPs). Previous work established exponential lower bounds in the number of latent contexts for the LMDP class. This puts forward a question: under which natural assumptions a near-optimal policy of an LMDP can be efficiently learned? In this work, we study the class of LMDPs with {\em prospective side information}, when an agent receives additional, weakly revealing, information on the latent context at the beginning of each episode. We show that, surprisingly, this problem is not captured by contemporary settings and algorithms designed for partially observed environments. We then establish that any sample efficient algorithm must suffer at least $\Omega(K^{2/3})$-regret, as opposed to standard $\Omega(\sqrt{K})$ lower bounds, and design an algorithm with a matching upper bound.