Authors:
Ryota Kubo
1
;
Fumito Uwano
2
and
Manabu Ohta
2
Affiliations:
1
School of Engineering, Okayama University, 3-1-1 Tsushima-naka, Kita-ku, Okayama, Japan
;
2
Faculty of Environmental, Life, Natural Science and Technology, Okayama University, 3-1-1 Tsushima-naka, Kita-ku, Okayama, Japan
Keyword(s):
Commonsense Knowledge, Reinforcement Learning, Deep Q-Network, Reward Design.
Abstract:
In text-based reinforcement learning, an agent learns from text to make appropriate choices, with a focus on addressing challenges associated with imparting commonsense knowledge to the learning agent. The commonsense knowledge requires the agent to understand not only the context but also the meaning of textual data. However, the methodology has not been established, that is, the effects on the agents, state-action space, reward, and environment that constitute reinforcement learning are not revealed. This paper focused on the reward for the commonsense knowledge to propose a new reward design method on the existing learning framework called ScriptWorld. The experimental results let us discuss the influence of the reward on the acquisition of commonsense knowledge by reinforcement learning.