Toward a Unified Benchmark and Taxonomy of Stochastic Environments
Aryan Amit Barsainyan, Jing Yu Lim, Dianbo Liu
公開日: 2025/9/1
Abstract
Reinforcement Learning (RL) agents have achieved strong results on benchmarks such as Atari100k, yet they remain limited in robustness to real-world conditions. Model-Based RL approaches that rely on learned World Models often struggle in environments with true stochasticity and partial observability, despite their theoretical grounding in POMDPs. Current benchmarks rarely capture these challenges, focusing instead on deterministic or overly simplified settings, and the lack of a clear taxonomy of stochasticity further hampers systematic evaluation. To address this gap, we introduce STORI (STOchastic-ataRI), a benchmark that incorporates diverse stochastic effects and enables rigorous assessment of RL methods under varied forms of uncertainty. In addition, we propose a taxonomy of stochasticity in RL environments, providing a unified framework for analyzing and comparing approaches.