Deep Reinforcement Learning for Traffic Signal Control Optimization

Palfi, Bogdan (2022) Deep Reinforcement Learning for Traffic Signal Control Optimization. Bachelor's Thesis, Artificial Intelligence.

Preview	Text bAI_2022_PalfiB.pdf Download (832kB) | Preview
	Text Toestemming.pdf Restricted to Registered users only Download (120kB)

Abstract

The increase in the number of personal vehicles per capita often leads to traffic congestion in highly urbanized areas. A solution to this problem consists of optimizing traffic signal control (TSC) policies using deep reinforcement learning (DRL). For this reason, the current thesis analyzed the TSC performance of two DRL algorithms introduced by Sabatelli et al. (2020), namely Deep Quality-Value (DQV) and Deep Quality-Value-Max (DQV-Max). The two algorithms were compared to classic DRL algorithms such as Deep Q-Network (DQN) and Double Deep Q-Network (DDQN), as well as to heuristic strategies such as Longest-Queue-First or Fixed-Timings. All algorithms were trained and tested in a simulated environment, created using Eclipse SUMO, which involved a single cross intersection with 3 lanes. The DRL agents were tasked with reducing the waiting time at the intersection by selecting which of the lanes were granted a green light. The results showed that all DRL agents achieved lower average waiting times than the heuristic strategies. Overall, the DQN and DQV-Max algorithms were more stable and kept waiting times lower than DDQN and DQV. The latter algorithms presented some instability, indicated by occasional waiting time spikes. While the environment was not complex enough to provide incentives for real-world deployment, the thesis acted as a proof of concept for the DQV and DQV-Max algorithms in the context of TSC.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Sabatelli, M.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	26 Jul 2022 09:06
Last Modified:	26 Jul 2022 09:06
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/28167

Actions (login required)

View Item