Record:   Prev Next
Author Li, Sa
Title Mobile robot obstacle avoidance using learning classifier systems
book jacket
Descript 125 p
Note Source: Dissertation Abstracts International, Volume: 68-10, Section: B, page: 6861
Thesis (Ph.D.)--University of Alberta (Canada), 2007
During the past few decades there has been much development of model-based approaches to autonomous mobile robot obstacle avoidance. Much of the work centre on navigation using artificial intelligence (AI) techniques. The goal of this thesis is to explore alternative approaches in the evolution of robot intelligence, rather than its prior design. It is proposed to use a method of evolutionary computing (EC)---the learning classifier system (LCS)---that possesses salient properties necessary for learning. Several LCS-type techniques are used to evolve obstacle-avoidance rules for mobile robot navigation. The robot learns these rules via feedback from the environment, available as sonar readings. Conventional LCSs (Michigan-style LCSs) show evidence of shortcomings: they become trapped in local minima, they lose desirable rules, and they favor rules that are too general. Enhancements proposed in this thesis overcome these limitations, and can thus handle more complex situations than existing devices. The inclusion of a dynamics feature helps to prevent oscillations of the learning process and makes robot learning converge quickly after each collision. To combine the structural learning ability of an LCS and the continuous adaptive characteristics of a neural network (NN), an LCS-like NN is constructed by mapping an LCS onto an NN. The system employs a genetic algorithm (GA) in a distinct way which demonstrates its co-adaptive characteristics. The LCS-like NN described in this thesis proves to be effective by evolving a diverse population of hidden nodes that compete to perform robot actions
School code: 0351
DDC
Host Item Dissertation Abstracts International 68-10B
Subject Engineering, Electronics and Electrical
Artificial Intelligence
Computer Science
0544
0800
0984
Alt Author University of Alberta (Canada)
Record:   Prev Next