Multi-Agent Reinforcement Learning Based on Dual Model Loss Prediction

In recent years, Multi-Agent Reinforcement Learning (MARL) has made breakthrough progress, demonstrating superior collaborative capabilities over human experts in complex scenarios and providing new solutions for multi-agent tasks. However, MARL faces significant challenges in practical deployment: the interaction between multiple agents and the environment reduces system stability, and the dimensions of the joint action space and global state space increase exponentially with the number of agents. This paper proposes a multi-agent reinforcement learning algorithm based on Dual Model Loss Prediction (DMLP). It combines state and action novelty to encourage agents to continuously explore and optimize the joint policy during training. The DMLP model consists of a feature extractor, a forward model, and a backward model. The feature extractor extracts state features; the forward model takes the current state features as input and predicts the next state features; the backward model takes the current and next state features as input to predict the agent’s action. Ultimately, DMLP fuses the state prediction error and action prediction error to form an intrinsic reward that drives the exploration of the joint policy. DMLP can be seamlessly integrated with existing multi-agent reinforcement learning algorithms. Experiments show that DMLP can effectively improve algorithm performance.