Policy Learning

Learning from a stream of experience as it arrives, also known as streaming learning, is a core part of natural learning. However, reliable streaming learning has remained a persistent challenge in modern deep reinforcement learning (RL). Instead, most deep RL algorithms learn from old experience by storing past interactions in a buffer. We show that both classical streaming RL, such as Q-learning and actor-critic, when used with deep neural networks, and batch deep RL, such as PPO, SAC, and DQN, when adapted to the streaming setting, often fail to learn. Across 58 Atari games and 50 continuous-control tasks, we find that these methods, in aggregate, perform close to random policies despite extensive task-specific hyperparameter searches. We call this pattern stream barrier. Here, we introduce Stream-X, a shared recipe for streaming deep RL algorithms that combines signal normalization, representation stabilization, and controlled parameter updates. By applying Stream-X to several base streaming RL algorithms, we provide the first family of deep RL algorithms to overcome the stream barrier. Using one prescribed hyperparameter configuration per algorithm across tasks, Stream-X substantially improves aggregate performance, often on par with batch RL algorithms. Beyond these benchmarks, we demonstrate learning with Stream-X algorithms under nonstationarity and resource constraints. Stream-AC, one of the Stream-X algorithms, repeatedly recovers performance across alternating floor-friction regimes in simulation, outperforming the evaluated PPO and SAC baselines. It also learns a heading tracking task on a robot using proprioceptive and visual features from the on-board camera in a naturally changing laboratory environment. Stream-Q learns a Pong game from pixels directly on an ESP32-S3 microcontroller, a device with limited compute and memory.

In the zero-shot policy transfer (ZSPT) setting for contextual Markov decision processes (CMDP), agents train on a fixed, finite set of contexts and must generalize to new ones. Recent work has demonstrated that training on additional states, even if they are irrelevant for solving the current context, can improve generalization to unseen contexts. In this paper, we demonstrate that training on these states can indeed improve generalization, but can come at a cost of reducing the accuracy of the learned value function, which should hurt generalization. We hypothesize and demonstrate that increasing the agent's coverage by training on these additional states while also increasing the accuracy improves generalization even further. Inspired by this, we propose a simple approach Explore-Go that leverages existing pure exploration strategies in a new way: by introducing a pure exploration phase at the start of each training episode. Unlike previous approaches that apply exploration strategies for the purpose of improving generalization, our approach can be combined with both on- and off-policy algorithms. We demonstrate the effectiveness of Explore-Go when combined with several popular algorithms and show an increase in test-time performance across several generalization benchmarks, even partially observable ones. With this, we hope to provide practitioners with a simple modification that can significantly improve the generalization of their agents.

Offline reinforcement learning enables evaluation and optimization of sequential decisions from historical data, when it is not possible to deploy new policies online due to safety, cost, and other concerns. Big data advances enable rich state information, but may naively include reward- and action- irrelevant dynamics that are ultimately unnecessary for learning optimal actions. We introduce state abstractions that target preservation of the difference-of-Q functions, and we propose to learn these abstractions via causal machine learning of the difference-of-Q function and standard statistical sparse learning. Under a nonparametric additive-rewards model, we characterize when decision-centered abstractions are simpler than the full state space, motivating our estimation procedure. We develop a dynamic generalization of the R learner (Nie et al. 2021, Lewis and Syrgkanis 2021) for estimating difference of Q-functions, for discrete-valued actions a, a0. We leverage orthogonal estimation to improve convergence rates, even if the required estimates of Q and behavior policy converge at slower rates and prove consistency of policy optimization under a margin condition. The method can leverage black-box estimators of the Q-function and behavior policy to target estimation of a more structured Q-function contrast, and uses simple squared-loss minimization. We demonstrate variance improvements from our estimator and how our approach enables us to isolate the information needed for sequential decision-making, which can be less than that for state prediction, in simulated data and simulator-augmented real data.

Adversarial Imitation Learning (AIL) faces challenges with sample inefficiency because of its reliance on sufficient on-policy data to evaluate the performance of the current policy during reward function updates. In this work, we study the convergence properties and sample complexity of off-policy AIL algorithms. We show that, even in the absence of importance sampling correction, reusing samples generated by the $o(\sqrt{K})$ most recent policies, where $K$ is the number of iterations of policy updates and reward updates, does not undermine the convergence guarantees of this class of algorithms. Furthermore, our results indicate that the distribution shift error induced by off-policy updates is dominated by the benefits of having more data available. This result provides theoretical support for the sample efficiency of off-policy AIL algorithms. To the best of our knowledge, this is the first work that provides theoretical guarantees for off-policy AIL algorithms.

Modern AI systems such as self-driving cars and game-playing agents can achieve superhuman performance, but often lack human-like generalization, interpretability, and inter-operability with human users. Inspired by the rich interactions between language and decision-making in humans, we introduce Policy Learning with a Language Bottleneck (PLLB), a framework enabling AI agents to generate linguistic rules that capture the high-level strategies underlying rewarding behaviors. PLLB alternates between a rule generation step guided by language models, and an update step where agents learn new policies guided by rules, even when a rule is insufficient to describe an entire complex policy. Across five diverse tasks, including a two-player signaling game, maze navigation, image reconstruction, and robot grasp planning, we show that PLLB agents are not only able to learn more interpretable and generalizable behaviors, but can also share the learned rules with human users, enabling more effective human-AI coordination. We provide source code for our experiments at https://github.com/meghabyte/bottleneck .

We consider the problem of dynamically selecting among competing strategies using observable predictors. To mitigate estimation noise and excessive switching, we regularise the estimated policy with a switching cost, which may exceed actual transaction costs when these are too small to provide sufficient regularisation. Since switching costs make current choices affect future decisions, we formulate the problem using dynamic programming. The optimal policy is estimated using a reinforcement learning algorithm based on iterated regressions, making the procedure simple to implement. We apply the method to monthly rotation among momentum, value, profitability, and defensive U.S. equity factors over a sample of about 60 years. Using a parsimonious set of state variables, the estimated policy outperforms both the myopic dynamic policy and the best fixed factor in hindsight in out-of-sample evaluation. Under weak conditions, we show theoretically that the decision rule estimated from empirical data is asymptotically consistent for the unknown optimal decision rule. Simulation results illustrate the finite-sample relevance of these asymptotic results.

We study formal policy synthesis for continuous-state stochastic systems under linear temporal logic specifications. The product of the system with the automaton of the specification has a hybrid state space with sparse rewards. We introduce a generalized optimal backup order, defined in reverse to a topological order over automaton states, that guides value backups and provably preserves optimality. We further present a model-free actor-critic algorithm whose policy evaluation solves a constrained optimization problem by the augmented Lagrangian method, yielding hyperparameter self-tuning, and prove its optimality and convergence in the tabular case. Since integer encodings of automaton states impose a spurious ordinal relationship on functions learned by one network, we dedicate a value and a policy network to each automaton state (modular learning). The algorithm matches or outperforms PPO, DQN, and A2C on CartPole, and on a Dubins car under a temporal specification the topological order and modular learning raise the success rate from 26.0% to 71.5%.

This paper studies experimental designs for estimation and inference on policies with spillover effects. Units are organized into a finite number of large clusters and interact in unknown ways within each cluster. First, we introduce a single-wave experiment that estimates the marginal effect of a change in treatment probabilities, taking spillover effects into account. Using the marginal effect, we propose a policy optimality test and improve the population's average outcome through marginal changes. Second, we design a multi-wave experiment to estimate optimal treatment rules, with strong regret guarantees. We provide an implementation in a large-scale field experiment studying climate adaptation.