Adversarial Controllers for Robust Evolutionary Control
Loading...
Authors
Gonzales, Sierra A.
Issue Date
2018
Type
Thesis
Language
Keywords
Adversarial Learning , Competitive Co-evolution , Evolutionary Algorithms , Min-max Control , Neural Networks , Simulation versus Reality Gap
Alternative Title
Abstract
Controllers in real-world applications often undergo challenges due to disturbances that cannot be approximated or modeled. A more robust controller can be created by dynamically changing the training environment through the use of a second entity acting on the system. The competition that occurs between two entities can generalize a solution to some of those unmodeled disturbances. In this work, two methods of training a baseline controller with an Antagonist are introduced in order to produce a controller more adaptable to environmental variances during testing. The first method introduced is the novel Min-Max Evolutionary Algorithm which produces two controllers in a co-evolutionary fashion: the Primary controller seeks to minimize error, while an Antagonist controller acts as a control input into the system to maximize the error. The second method produces a Primary controller that minimizes error while an Antagonist, that manipulates system variables like starting states and goal values, to create disturbances with the objective of maximizing error. This research uses a mass-spring-damper control domain to show that the way in which the Antagonist is implemented in training affects the robustness of the final performance compared to the non-trained Primary controller in a test environment.