Moving object recognition and avoidance using reinforcement learning

Cosma, Radu Alexandru (2021) Moving object recognition and avoidance using reinforcement learning. Bachelor's Thesis, Artificial Intelligence.

Preview	Text Bachelor_s_Project.pdf Download (2MB) | Preview
	Text Toestemming.pdf Restricted to Registered users only Download (118kB)

Abstract

This thesis explores the best way to handle multiple consecutive opponents in a reinforcement learning context. Q-learning, Double Q-learning and Sarsa, all using function approximation, are compared on a Gridworld game task with multiple levels in which an agent needs to pass an opponent and reach the goal. Monte-Carlo rollouts are used to improve action selection together with opponent models, represented using Multi-Layer Perceptrons. A comparison is made between different opponent modelling setups, with one opponent model for all opponents and two novel techniques compared. The novel techniques involve recognising whether an opponent has been previously encountered or is a new opponent. This is done by comparing the prediction losses on trajectories on which an opponent model was trained with the prediction losses on the new opponent. One method checks if these losses come from the same distribution, while the other whether there was a change-point in the losses when represented as a time series. Results show that the novel methods can predict the opponent considerably better on an illustrative example, with a more reduced improvement on more general deterministic opponents and no improvement on opponents with randomness. Overall mean final rewards are similar regardless of opponent modelling technique, with Sarsa performing best.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Wiering, M.A.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	10 Aug 2021 09:04
Last Modified:	10 Aug 2021 09:04
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/25605

Actions (login required)

View Item