Deep reinforcement learning (DRL) achieves remarkable success in complex tasks, but its large model size makes it a major challenge to deploy it on resource-constrained platforms while maintaining high performance and achieving high compression ratio. In this paper, we propose a DRL model compression method that combines group regularization pruning and random sketches. We first verify the effectiveness in a classic DRL environment and then demonstrate its application value in a multi-agent confrontation environment. Experimental results show that our method achieves a high compression ratio and superior policy performance over baselines.