Date of Award

5-31-2020

Document Type

Thesis (Undergraduate)

Department or Program

Department of Computer Science

First Advisor

Richard Granger

Second Advisor

Venkatramanan Subrahmanian

Abstract

This study explores how various constraints on a computer agent's memory and recall capacities affect how it performs a simple reinforcement learning task: the card-matching memory game "Concentration". Existing computer agents can solve this task easily, but humans struggle with it, even though its rules and objectives are simple. Why is this the case? We identify specific human memory limitations that may be at play: decaying of memories over time and remembering broad characteristics of card locations and faces while forgetting card specifics. Through building and testing a reinforcement learning agent with these human-like memory constraints, we find that they each lead to humanlike restrictions on agent performance. This work contributes to questions of comparison between computer and human memory capabilities, to questions of designing memory-augmented reinforcement learning systems, and to questions of aids for humans during memory-intensive tasks.

Comments

Originally posted in the Dartmouth College Computer Science Technical Report Series, number TR2020-886.

Download
COinS