Improving Learning Stability for Reinforcement Learning Agent

We present a method, which organically combines the actor/critic architecture with the self-organizing feature map (SOFM), and the results of its research aimed at improving the learning stability for reinforcement learning agent. A model is proposed based on the SOFM that receives as input the continuous state space and produces as output neurons, which then are mapped to BOXES. Our model extends the actor/critic architecture so that inactive BOXES may learn appropriate eligibility trace from active BOXES in order to improve learning stability for reinforcement learning agent. Experimental results obtained from a simulation show that our model is capable of learning a useful partitioning of the continuous state space and improving learning stability for reinforcement learning agents