Attention-Based Mechanisms for Cognitive Reinforcement Learning

Abstract

In this paper, we propose a cognitive reinforcement learning method based on an attention mechanism (CRL-CBAM) to address the problems of complex interactive communication, limited range, and time-varying communication topology in multi-intelligence collaborative work. The method not only combines the efficient decision-making capability of reinforcement learning, the representational capability of deep learning, and the self-learning capability of cognitive learning but also inserts a convolutional block attention module to increase the representational capability by using the attention mechanism to focus on important features and suppress unnecessary ones. The use of two modules, channel and spatial axis, to emphasize meaningful features in the two main dimensions can effectively aid the flow of information in the network. Results from simulation experiments show that the method has more rewards and is more efficient than other methods in formation control, which means a greater advantage when dealing with scenarios with a large number of agents. In group containment, the agents learn to sacrifice individual rewards to maximize group rewards. All tasks are successfully completed, even if the simulation scenario changes from the training scenario. The method can therefore be applied to new environments with effectiveness and robustness.