Online and offline training for adaptive neuro-fuzzy inference systems using deep and reinforcement learning with hierarchical rule-base reduction

Ahn, Woojin

Permalink

https://hdl.handle.net/2142/115470 Copy

Description

Title

Online and offline training for adaptive neuro-fuzzy inference systems using deep and reinforcement learning with hierarchical rule-base reduction

Author(s)

Ahn, Woojin

Issue Date

2022-04-25

Director of Research (if dissertation) or Advisor (if thesis)

Norris, William

Department of Study

Industrial&Enterprise Sys Eng

Discipline

Systems & Entrepreneurial Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Autonomous Vehicle
• Adaptive Neuro-Fuzzy Inference System
• Reinforcement Learning
• Deep Deterministic Policy Gradient
• Fuzzy Controller
• Artificial Intelligence
• Hierarchical Rule Base Reduction

Abstract

This study successfully implemented an Adaptive Neuro-Fuzzy Inference System (ANFIS) [1] vehicle controller trained online and offline with machine learning, deep learning, and reinforcement learning. It was applied to an autonomous skid steering off-road robot path tracking control, as one of the potential applications for this approach. The ANFIS controller was a fuzzy system transformed into a neural network structure to self train. The fuzzy system is explainable because it uses linguistic variables with a logical rule-base, and the neural network is trainable and directly transforms from the fuzzy system structure. The ANFIS, as an explainable artificial intelligence, is designed as a fuzzy logic based human decision-making model (HDMM) with Fuzzy Relations Control Strategy (FRCS) [2] to dramatically reduce computational time and leverage the advantages of both the fuzzy system and neural network. The ANFIS controller was trained using a dataset collected from the expert system in simulation with offline supervised learning. The controller replicated and improved the behavior of the expert model after the offline training. Also, the ANFIS controller was trained using online reinforcement learning on the actual vehicle while driving, which enabled the controller to train itself without any datasets. The result of the supervised learning showed that the error between the ANFIS controller and the expert system was 9.28%. The result of the ANFIS controller trained using online reinforcement learning showed that the trained ANFIS controller performed over 87% in simulation and 73% on the actual vehicle better than the untrained ANFIS controller on five different test courses.

Graduation Semester

2022-05

Type of Resource

Thesis

Copyright and License Information

Copyright 2022 Woojin Ahn

Owning Collections