TABS: Strategic Game-Based Multi-Stage Reinforcement Learning Challenge

Hayeong Lee, JunHyeok Oh, Byung-Jun Lee

Keywords: Reinforcement Learning, Multi Agent Reinforcement Learning, JAX

Abstract: The design of environments plays a critical role in shaping the development and evaluation of reinforcement learning (RL) algorithms. While existing benchmarks have supported significant progress in both single-agent and multi-agent settings, many real-world systems involve multi-stage structures with tightly coupled decision points at each stage. These settings require agents not only to perform well within each stage but also to coordinate effectively across them. We introduce the Totally Accelerated Battle Simulator (TABS), a complex multi-stage environment suite implemented in JAX to enable accelerated training and scalable experimentation. Each TABS task consists of sequential stages with interdependencies, where only the output of one stage is forwarded to the next. This multi-stage structure makes effective exploration challenging, often steering agents toward locally optimal behaviors that limit overall performance. Our empirical analysis shows that standard RL baselines struggle to solve TABS tasks, illustrating the difficulty of learning coherent strategies across interdependent stages. TABS provides a controlled and extensible framework for studying multi-stage decision-making challenges, supporting future research on RL methods capable of operating effectively in structured domains. Our code is available at:~\url{https://anonymous.4open.science/r/TABS-0E4B}.

Primary Area: datasets and benchmarks

Submission Number: 3707