MEAL: A Benchmark for Continual Multi-Agent Reinforcement Learning

The article introduces MEAL (Multi-agent Environments for Adaptive Learning), a benchmark specifically designed for continual multi-agent reinforcement learning (RL). Utilizing JAX and GPU acceleration, MEAL allows for the training of RL agents on sequences of up to 100 tasks within a few hours on a single GPU, addressing the limitations of existing benchmarks that typically focus on only 3-10 tasks. This development is significant for practitioners as it uncovers failure modes in long task sequences that are not observable in shorter ones, thus providing deeper insights for designing robust multi-agent systems.