Safe offline reinforcement learning aims to learn policies that maximize cumulative rewards while adhering to safety constraints, using only offline data for training. A key challenge is balancing safety and performance, particularly when the policy encounters out-of-distribution (OOD) states and actions, which can lead to safety violations or overly conservative behavior during deployment. To address these challenges, we introduce Feasibility Informed Advantage Weighted Actor-Critic (FAWAC), a method that prioritizes persistent safety in constrained Markov decision processes (CMDPs). FAWAC formulates policy optimization with feasibility conditions derived specifically for offline datasets, enabling safe policy updates in non-parametric policy space, followed by projection into parametric space for constrained actor training. By incorporating a cost-advantage term into Advantage Weighted Regression (AWR), FAWAC ensures that the safety constraints are respected while maximizing performance. Additionally, we propose a strategy to address a more challenging class of problems that involves tempting datasets where trajectories are predominantly high-rewarded but unsafe. Empirical evaluations on standard benchmarks demonstrate that FAWAC achieves strong results, effectively balancing safety and performance in learning policies from the static datasets.

In safe offline reinforcement learning (RL), the objective is to develop a policy that maximizes cumulative rewards while strictly adhering to safety constraints, utilizing only offline data. Traditional methods often face difficulties in balancing these constraints, leading to either diminished performance or increased safety risks. We address these issues with a novel approach that begins by learning a conservatively safe policy through the use of Conditional Variational Autoencoders, which model the latent safety constraints. Subsequently, we frame this as a Constrained Reward-Return Maximization problem, wherein the policy aims to optimize rewards while complying with the inferred latent safety constraints. This is achieved by training an encoder with a reward-Advantage Weighted Regression objective within the latent constraint space. Our methodology is supported by theoretical analysis, including bounds on policy performance and sample complexity. Extensive empirical evaluation on benchmark datasets, including challenging autonomous driving scenarios, demonstrates that our approach not only maintains safety compliance but also excels in cumulative reward optimization, surpassing existing methods. Additional visualizations provide further insights into the effectiveness and underlying mechanisms of our approach. Although Reinforcement learning (RL) is a popular approach for decision-making and control applications across various domains, its deployment in industrial contexts is limited by safety concerns during the training phase. In traditional online RL, agents learn optimal policies through trial and error, interacting with their environments to maximize cumulative rewards. This process inherently involves exploration, which can lead to the agent encountering unsafe states and/or taking unsafe actions, posing substantial risks in industrial applications such as autonomous driving, robotics, and manufacturing systems (Garcıa & Fernández, 2015; Gu et al., 2022; Moldovan & Abbeel, 2012; Shen et al., 2014; Yang et al., 2020). The primary challenge lies in ensuring that the agent's learning process does not compromise safety, as failures during training can result in costly damages, operational disruptions, or even endanger human lives (Achiam et al., 2017; Stooke et al., 2020). To address these challenges, researchers have explored several approaches aimed at minimizing safety risks while maintaining the efficacy of RL algorithms. One effective method to mitigate safety risks associated with training an agent is offline RL. This dataset comprises trajectory rollouts generated by an arbitrary behavior policy or multiple policies, collected beforehand.

The study proposes the reformulation of offline reinforcement learning as a regression problem that can be solved with decision trees. Aiming to predict actions based on input states, return-to-go (RTG), and timestep information, we observe that with gradient-boosted trees, the agent training and inference are very fast, the former taking less than a minute. Despite the simplification inherent in this reformulated problem, our agent demonstrates performance that is at least on par with established methods. This assertion is validated by testing it across standard datasets associated with D4RL Gym-MuJoCo tasks. We further discuss the agent's ability to generalize by testing it on two extreme cases, how it learns to model the return distributions effectively even with highly skewed expert datasets, and how it exhibits robust performance in scenarios with sparse/delayed rewards.