Investigating Comparisons Between Human Learning and Machine Learning Using the Dots and Boxes Game

Abstract

From its origins, machine learning has drawn inspiration from the human brain and its thought processes. Despite machine learning drawing so much from human learning, comparisons of the progression of learning between the two are fraught with difficulties. This study aims to explore the similarities and differences between human and machine learning, using a perfect information board game, Dots and Boxes, that both humans and machine learning agents learn to play by training against the same ‘box-greedy’ policy agent. Three q-learning reinforcement learning agents have been created with three distinct levels of strategy to compare against the learning progression and strategy of 62 human volunteers developed over 20 games each. Volunteers were also asked to complete BIS/BAS and CRT-MQC4 questionnaires after completing the Dots and Boxes task to study behavioural metrics that affect decision making, such as system 1 & 2 thinking and sensitivity to reward or punishment. Results from comparisons show how important context and prior knowledge is to human learning, and how machine learning agents can generalise better and reach optimal strategy faster with prior knowledge. Additional observations around how behavioural metrics affect individual decision making and correlations between machine learning agent strategy and human participant data allow for further comparisons to be made between human and machine learning.