T. Mao, X. Sun, and L.E. Ray (USA)
Abstract state, Multi-robot systems, Role selection, and einforcement learning.
Although reinforcement learning using Q-learning has been successfully applied to multi-robot systems, the technique still has computational and practical issues, specifically, representation and exploration of large learning spaces. This paper uses state abstraction to reduce the size of the state space and minimize the action space. Abstract state space representation requires less memory to store Q values and increases numerical efficiency of visiting possible states. Simulation results within the context of a multi-robot foraging problem demonstrate that abstract state-based reinforcement learning converges to satisficing policies for role emergence and task execution in a shorter time, compared to performance of reinforcement learning for the same task but with a detailed representation of the state space.
Important Links:
Go Back
