Offline Pre-trained Multi-agent Decision Transformer

Abstract Offline reinforcement learning leverages previously collected offline datasets to learn optimal policies with no necessity access the real environment. Such a paradigm is also desirable for multi-agent (MARL) tasks, given combinatorially increased interactions among agents and However, in MARL, of pre-training online fine-tuning has not been studied, nor even or benchmarks MARL research are available. In this paper, we facilitate by providing large-scale using them examine usage decision transformer context MARL. We investigate generalization following three aspects: 1) between single multiple agents, 2) from pretraining fine tuning, 3) that downstream tasks few-shot zero-shot capabilities. start introducing first dataset diverse quality levels based on StarCraftII environment, then propose novel architecture (MADT) effective learning. MADT transformer’s modelling ability sequence integrates it seamlessly both tasks. A significant benefit learns generalizable can transfer different types under task scenarios. On StarCraft II dataset, outperforms state-of-the-art (RL) baselines, including BCQ CQL. When applied pre-trained significantly improves sample efficiency enjoys strong performance few-short cases. To best our knowledge, work studies demonstrates effectiveness models terms generalizability enhancements