Combining affective intelligence with learning to improve action selection in decision-making agents

Article type: Research Article

Authors: Williams, Jason B. | Zhang, Xiaoqin^*

Affiliations: Department of Computer and Information Science, University of Massachusetts, Dartmouth, MA, USA

Correspondence: [*] Corresponding author: Xiaoqin Zhang, Department of Computer and Information Science, University of Massachusetts, Dartmouth, MA, USA. E-mail: [email protected].

Abstract: Complex environments contain more information than either natural or artificial agents can fully process in a timely manner. Studies in neuroscience have demonstrated that natural agents utilize affect (or emotion) to filter out irrelevant inputs. In this work, we propose to integrate an affect filtering mechanism in artificial agents to improve the deliberation time for action selection in environment containing a massive number of selection options. To evaluate this model, we create two agent architectures: the first architecture is based on an active reinforcement learning algorithm and the second architecture utilizes a hybrid design with both active reinforcement learning and the affect-based filtering mechanism. We have compared the deliberation time and the overall utility score of these two agents in the same environment. The results showed that the affect-based filtering mechanism is effective in decreasing the deliberation time without compromising the agent’s utility score. The results from this study strengthen the premise that affect plays an important role in intelligent behavior.

Keywords: Affective-computing, decision-making, agent architecture, machine learning, reinforcement learning

DOI: 10.3233/HIS-180259

Journal: International Journal of Hybrid Intelligent Systems, vol. 15, no. 1, pp. 27-53, 2019