Year: 2020
Type: dissertation
Abstract: This thesis explores the field of \\emph{Affective Computing} and emotional intelligent machines, by implementing and investigating a specific approach for training emotional reinforcement learning agents, in the novel scenario of playing the video game Super Mario Bros.. Our approach, which is inspired by previous work in the field, is to collect specific types of data, from real humans playing the game, that correlates to emotional reactions, and then use this data to train a convolutional neural network (CNN) to predict this emotional reaction on the basis of information from the game environment. This CNN is then applied to create a custom reward scheme for a reinforcement learning agent, which introduces an \\emph{intrinsic} emotional reward, that is integrated with the pre-existing extrinsic reward of a standard deep Q network (DQN) algorithm. This thesis documents the work of (i) collecting and processing the required data into an open dataset called Toadstool, (ii) training CNN models on blood volume pulse data, to predict an emotional reaction based on frames from the game, (iii) the development of a DQN that learns to play the game and, (iv) the integration of the predicted emotional signal into the reward function of the DQN, to create an emotional DQN (EDQN). We compare the performance of various EDQN models with the standard DQN, and produce results that show that some implementations of the EDQN outperforms the DQN in certain respects.
Author: Henrik Svoren
Language: English
Topic: Emotion and Mood Recognition
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W3173742446