Application of action prediction in multi-robot reinforcement learning cooperation
Ningning Zhu · Computer Engineering and Applications Journal · 2013
In multi-robot systems, the spatial scale of reinforcement learning of the cooperation environment exploration is made up of the exponential function of the number of robots. And the enormous learning space results in the slow convergence rate. To solve this problem, a prediction-based reinforcement learning algorithm and the action selection strategy are applied to the research on multi-robot cooperation. By predicting the probability of actions that other robots may execute, the convergence rate of this algorithm is accelerated. The experimental results show that reinforcement learning algorithm based-on action prediction can achieve the multi-robot’s cooperation strategy much faster, compared to the primitive algorithm.