This paper introduces ASALT, a method for lateral transfer learning in multi-agent reinforcement learning that accommodates mismatched state-space dimensionalities between source and target domains. The approach uses observation-level and state-level adapters to map inputs into a shared embedding space, enabling effective knowledge transfer across heterogeneous environments.
- ASALT incorporates both observation-level and state-level adapters to handle differing observation and global state spaces.
- Experimental results show superior sample efficiency and global return in cooperative settings compared to existing baselines.
- The method mitigates negative transfer, a common obstacle when transferring policies between domains with different dimensions.
- Effectiveness depends on the degree of mismatch between source and target domains.
ASALT addresses the limitation of prior MARL transfer approaches that require identical dimensionalities, allowing for more flexible strategy transfer across diverse multi-agent systems.