Reinforcement Learning with Adaptive Networks

The reinforcement learning (RL) with the incremental normalized Gaussian networks (INGnet) using the state list was proposed in order to improve the learning efficiency. This learning system was applied to the garage parking control, the walking control of the stilt-type biped robot on the slope, and the wall avoidance control of Khepera robot. Since in the ordinary condition to add the new processing units they were not set in the central region of the garage parking field or around the states during the normal walking, RL with INGnet did not learn the control effectively. RL with INGnet using the state list set the processing units at the necessary positions and learned the garage parking control and the walking control of the stilt-type biped robot on the slope. However, since it is easy to learn the wall avoidance control, the state list was not needed for RL with INGnet: RL with INGnet without the state list was able to learn it.