[2I5-GS-2] Machine learning: Cognition and decision support

[2I5-GS-2-03] Cognitive Satisfaction Exploration Method Based on Approximation Error in Real World Data

〇Akane Minami¹, Yukina Kobayashi¹, Yu Kono¹, Tatsuji Takahashi¹ (1. School of Science and Engineering, Tokyo Denki University)

Keywords:Reinforcement learning, Machine learning, Contextual Bandit Problem, Satisficing

Reinforcement learning has a search concept that requires data collection by itself, and the optimal search algorithm has been clarified for the bandit problem that is one of simple reinforcement learning. However, It can’t be guaranteed for the contextual bandit problem using function approximation. Therefore, we verified a search algorithm different from the conventional one. It’s known that humans have a target level of reward and quickly search skills that reach the target level (satisfaction). Linear Risk-sensitive Satisficing (LinRS) that is adapted to the contextual bandit problem that applies this satisfaction, has achieved better results compared to the existing algorithm using artificial distribution. We verified the contextual bandit problem using the real data. It’s said that the performance of the real data is worse than the artificial one, and we describe about how to adjust the target level for appropriate search using LinRS as countermeasures.