The limited generalization ability of reinforcement learning
constrains its potential applications, particularly in complex scenarios
such as multi-agent systems. To overcome this limitation and enhance
the generalization capability of MARL algorithms, this paper proposes
a three-stage method that integrates domain randomization and domain
adaptation to extract effective features for policy learning. Specifically,
the first stage samples environments provided for training and testing
in the following stages using domain randomization. The second stage
pretrains a domain-invariant feature extractor (DIFE) which employs
cycle consistency to disentangle domain-invariant and domain-specific
features. The third stage utilizes DIFE for policy learning. Experimental
results in MPE tasks demonstrate that our approach yields better performance
and generalization ability. Meanwhile, the features captured by
DIFE are more interpretable for subsequent policy learning in visualization
analysis.