IISER Kolkata ePrints Repository

Reinforcement Learning to Train Agents to Play Games

Lamba, Vishal (2019) Reinforcement Learning to Train Agents to Play Games. Masters thesis, Indian Institute Of Science Education and Research Kolkata.

[img] PDF (MS dissertation of Vishal Lamba (14MS172)) - Submitted Version
Restricted to Repository staff only

Download (529Kb)

    Abstract

    This thesis focuses on the challenge of decoupling state perception and function approx- imation when applying Deep Learning Methods within Reinforcement Learning. As a starting point, high-dimensional states were considered, this being the fundamental lim- itation when applying Reinforcement Learning to real world tasks. Different approaches to solve the problem of estimation Q-values of states in an en- vironment are discussed. The key idea is to estimating Q-values for large state-spaces, where tabular approaches are no longer feasible. As a mean to perform Q-function ap- proximation, we search for supervised learning methods within Deep Learning. Using python, a model for the environment of the game of 'Pong' was created and using Deep Q-Learning an agent was trained to achieve excellent performance at playing the game.

    Item Type: Thesis (Masters)
    Additional Information: Supervisor: Prof. P.K. Panigrahi
    Uncontrolled Keywords: Deep Learning Methods; Deep Neural Networks; Function Approximation; Reinforcement Learning; State Perception
    Subjects: Q Science > QA Mathematics
    Divisions: Department of Mathematics and Statistics
    Depositing User: IISER Kolkata Librarian
    Date Deposited: 09 Oct 2019 14:35
    Last Modified: 09 Oct 2019 14:35
    URI: http://eprints.iiserkol.ac.in/id/eprint/861

    Actions (login required)

    View Item