Home // International Journal On Advances in Intelligent Systems, volume 16, numbers 3 and 4, 2023 // View article

Comparing Concept Acquisition in Human & Superhuman DeepRL Models

Authors:
Anthony Marchiafava
Atriya Sen

Keywords: Artificial Intelligence; Deep Reinforcement Learning; Reinforcement Learning; Deep Learning; Explainable AI

Abstract:
In order to better exploit Deep Reinforcement Learning (DeepRL) systems such as DeepMind’s AlphaGo & AlphaZero, it is desirable to understand how they acquire knowledge, and how human knowledge acquisition can contribute to or benefit from such an understanding. We analyze a series of DeepRL models called Maia, trained to play the board game of chess in a human--like fashion, to study if these models acquire concepts differently from self--trained DeepRL models such as AlphaZero. Our results indicate that human chess players may acquire concepts differently from self--trained models. We further discuss some of the potential consequences of such an outcome, and opportunities for future work.

Pages: 19 to 30

Copyright: Copyright (c) to authors, 2023. Used with permission.

Publication date: December 30, 2023

Published in: journal

ISSN: 1942-2679