[3Xin4] Poster session 1

[3Xin4-26] Novel reinforcement learning model by tiled neural network based on the cortico-basal ganglia circuit

〇Tetsuji Kita¹, Shunya Onoi¹, Ryo Takehara¹, Kojiro Hirokane^1,2, Takashi Kitsukawa¹ (1.Ritsumeikan University, 2.Osaka University)

Keywords:reinforcement learning, basal ganglia, neural network

Artificial reinforcement learning algorithms have evolved in the direction of practicality, many of which take advantage of reward prediction errors. Animals also can perform reinforcement learning, which means that animals have neural circuits engaged in the reinforcement learning. In fact, dopamine neurons in the substantia nigra were found to represent the reward prediction error suggesting that neural circuits in the basal ganglia, which are the targets of the dopamine neurons, may be involved in the information processing related to the reinforcement learning. However, it remains to reveal reinforcement learning algorithms at the neural circuit level. In this study, we constructed a neural network model for reinforcement learning by parallelly tiling of neural circuits of the cortex-basal ganglia using reward prediction errors from dopamine neurons. We compared our model to the conventional TD learning using Markov decision process tasks, Maze and CartPole. As a result, we found that our model reproduced the activity patterns of dopamine neurons as observed in animals performing reinforcement learning, and demonstrated performance comparable to that of TD learning.