Work

Automated reverse engineering of agent behaviors

Year: 2016
Type: dissertation
Abstract: This thesis concerns the automated reverse engineering of agent behaviors.It proposes a metric-free coevolutionary approach-Turing Learning, which allows a machine to infer the behaviors of agents (simulated or physical ones), in a fully automated way.Turing Learning consists of two populations.A population of models competitively coevolves with a population of classifiers.The classifiers observe the models and agents.The fitness of the classifiers depends solely on their ability to distinguish between them.The models, on the other hand, are evolved to mimic the behavior of the agents and mislead the judgment of the classifiers.The fitness of the models depends solely on their ability to 'trick' the classifiers into categorizing them as agents.Unlike other methods for system identification, Turing Learning does not require any predefined metrics to quantitatively measure the difference between the models and agents.The merits of Turing Learning are demonstrated using three case studies.In the first case study, a machine automatically infers the behavioral rules of a group of homogeneous agents through observation.A replica, which resembles the agents under investigation in terms of behavioral capabilities, is mixed into the group.The models are executed on the replica.This case study is conducted with swarms of both simulated and physical robots.In the second and third case studies, Turing Learning is applied to infer deterministic and stochastic behaviors of a single agent through controlled interaction, respectively.In particular, the machine is able to modify the environmental stimuli that the agent responds to.In the case study of inferring deterministic behavior, the machine can construct static patterns of stimuli that facilitate the learning process.In the case study of inferring stochastic behavior, the machine needs to interact with the agent on the fly through dynamically changing patterns of stimuli.This allows the machine to explore the agent's hidden information and thus reveal its entire behavioral repertoire.This interactive approach proves superior to learning only through observation.iii I consider the process of pursuing my PhD as a journey to learn not only about science, but also about life.Firstly, I would like to thank my supervisors, Dr. Roderich Groß and Prof. Stephen A. Billings for their support in this journey.The special thanks would be given to Roderich.He helped me from toddling to walking on the pathway in this research field by motivating, teaching and passing his professional experience
Author: Wei Li
Language: English
Open Access status: green
OpenAlex ID: W2409452136
Cites: 152
Cited by: 1
Related to: 10