We make two surprising observations.

While deep reinforcement learning has successfully solved many challenging control tasks, its real-world applicability has been limited by the inability to ensure the safety of learned policies. We propose an approach to verifiable reinforcement learning by training decision tree policies, which can represent complex policies (since they are nonparametric), yet can be efficiently verified using existing techniques (since they are highly structured). The challenge is that decision tree policies are difficult to train. We propose VIPER, an algorithm that combines ideas from model compression and imitation learning to learn decision tree policies guided by a DNN policy (called the oracle) and its Q-function, and show that it substantially outperforms two baselines. We use VIPER to (i) learn a provably robust decision tree policy for a variant of Atari Pong with a symbolic state space, (ii) learn a decision tree policy for a toy game based on Pong that provably never loses, and (iii) learn a provably stable decision tree policy for cart-pole. In each case, the decision tree policy achieves performance equal to that of the original DNN policy.

The ability to learn from large batches of autonomously collected data for policy improvement -- a paradigm we refer to as batch online reinforcement learning -- holds the promise of enabling truly scalable robot learning by significantly reducing the need for human effort of data collection while getting benefits from self-improvement. Yet, despite the promise of this paradigm, it remains challenging to achieve due to algorithms not being able to learn effectively from the autonomous data. For example, prior works have applied imitation learning and filtered imitation learning methods to the batch online RL problem, but these algorithms often fail to efficiently improve from the autonomously collected data or converge quickly to a suboptimal point. This raises the question of what matters for effective batch online RL in robotics. Motivated by this question, we perform a systematic empirical study of three axes -- (i) algorithm class, (ii) policy extraction methods, and (iii) policy expressivity -- and analyze how these axes affect performance and scaling with the amount of autonomous data. Through our analysis, we make several observations. First, we observe that the use of Q-functions to guide batch online RL significantly improves performance over imitation-based methods. Building on this, we show that an implicit method of policy extraction -- via choosing the best action in the distribution of the policy -- is necessary over traditional policy extraction methods from offline RL. Next, we show that an expressive policy class is preferred over less expressive policy classes. Based on this analysis, we propose a general recipe for effective batch online RL. We then show a simple addition to the recipe of using temporally-correlated noise to obtain more diversity results in further performance gains. Our recipe obtains significantly better performance and scaling compared to prior methods.

Post rebuttal Thank the authors for the clarification. One minor point I realised is the equation between line 144 and 145. Is this constraint really a disjunction over partitions? If there is at least one partition the given state doesn't belong to, it would be always true because at least one of inner propositions will be true, wouldn't it? The trained decision tree policy allows for its verification in terms of, more specifically, correctness, stability and robustness.

While imitation learning requires access to high-quality data, offline reinforcement learning (RL) should, in principle, perform similarly or better with substantially lower data quality by using a value function. However, current results indicate that offline RL often performs worse than imitation learning, and it is often unclear what holds back the performance of offline RL. Motivated by this observation, we aim to understand the bottlenecks in current offline RL algorithms. While poor performance of offline RL is typically attributed to an imperfect value function, we ask: is the main bottleneck of offline RL indeed in learning the value function, or something else? To answer this question, we perform a systematic empirical study of (1) value learning, (2) policy extraction, and (3) policy generalization in offline RL problems, analyzing how these components affect performance. We make two surprising observations. First, we find that the choice of a policy extraction algorithm significantly affects the performance and scalability of offline RL, often more so than the value learning objective. For instance, we show that common value-weighted behavioral cloning objectives (e.g., AWR) do not fully leverage the learned value function, and switching to behavior-constrained policy gradient objectives (e.g., DDPG+BC) often leads to substantial improvements in performance and scalability. Second, we find that a big barrier to improving offline RL performance is often imperfect policy generalization on test-time states out of the support of the training data, rather than policy learning on in-distribution states. We then show that the use of suboptimal but high-coverage data or test-time policy training techniques can address this generalization issue in practice. Specifically, we propose two simple test-time policy improvement methods and show that these methods lead to better performance.

Implicit Q-learning (IQL) serves as a strong baseline for offline RL, which learns the value function using only dataset actions through quantile regression. However, it is unclear how to recover the implicit policy from the learned implicit Q-function and why IQL can utilize weighted regression for policy extraction. IDQL reinterprets IQL as an actor-critic method and gets weights of implicit policy, however, this weight only holds for the optimal value function. In this work, we introduce a different way to solve the implicit policy-finding problem (IPF) by formulating this problem as an optimization problem. Based on this optimization problem, we further propose two practical algorithms AlignIQL and AlignIQL-hard, which inherit the advantages of decoupling actor from critic in IQL and provide insights into why IQL can use weighted regression for policy extraction. Compared with IQL and IDQL, we find our method keeps the simplicity of IQL and solves the implicit policy-finding problem. Experimental results on D4RL datasets show that our method achieves competitive or superior results compared with other SOTA offline RL methods. Especially in complex sparse reward tasks like Antmaze and Adroit, our method outperforms IQL and IDQL by a significant margin.

Deep Reinforcement Learning (DRL) has achieved remarkable success in sequential decision-making problems. However, existing DRL agents make decisions in an opaque fashion, hindering the user from establishing trust and scrutinizing weaknesses of the agents. While recent research has developed Interpretable Policy Extraction (IPE) methods for explaining how an agent takes actions, their explanations are often inconsistent with the agent's behavior and thus, frequently fail to explain. To tackle this issue, we propose a novel method, Fidelity-Induced Policy Extraction (FIPE). Specifically, we start by analyzing the optimization mechanism of existing IPE methods, elaborating on the issue of ignoring consistency while increasing cumulative rewards. We then design a fidelity-induced mechanism by integrate a fidelity measurement into the reinforcement learning feedback. We conduct experiments in the complex control environment of StarCraft II, an arena typically avoided by current IPE methods. The experiment results demonstrate that FIPE outperforms the baselines in terms of interaction performance and consistency, meanwhile easy to understand.

In many real world applications, reinforcement learning agents have to optimize multiple objectives while following certain rules or satisfying a list of constraints. Classical methods based on reward shaping, i.e. a weighted combination of different objectives in the reward signal, or Lagrangian methods, including constraints in the loss function, have no guarantees that the agent satisfies the constraints at all points in time and lack in interpretability. When a discrete policy is extracted from an action-value function, safe actions can be ensured by restricting the action space at maximization, but can lead to sub-optimal solutions among feasible alternatives. In this work, we propose Multi Time-scale Constrained DQN, a novel algorithm restricting the action space directly in the Q-update to learn the optimal Q-function for the constrained MDP and the corresponding safe policy. In addition to single-step constraints referring only to the next action, we introduce a formulation for approximate multi-step constraints under the current target policy based on truncated value-functions to enhance interpretability. We compare our algorithm to reward shaping and Lagrangian methods in the application of high-level decision making in autonomous driving, considering constraints for safety, keeping right and comfort. We train our agent in the open-source simulator SUMO and on the real HighD data set.

While deep reinforcement learning has successfully solved many challenging control tasks, its real-world applicability has been limited by the inability to ensure the safety of learned policies. We propose an approach to verifiable reinforcement learning by training decision tree policies, which can represent complex policies (since they are nonparametric), yet can be efficiently verified using existing techniques (since they are highly structured). The challenge is that decision tree policies are difficult to train. We propose VIPER, an algorithm that combines ideas from model compression and imitation learning to learn decision tree policies guided by a DNN policy (called the oracle) and its Q-function, and show that it substantially outperforms two baselines. We use VIPER to (i) learn a provably robust decision tree policy for a variant of Atari Pong with a symbolic state space, (ii) learn a decision tree policy for a toy game based on Pong that provably never loses, and (iii) learn a provably stable decision tree policy for cart-pole.