Year: 2021
Type: dissertation
Abstract: Human-harming behaviours from Artificial Intelligence (AI) have long been a featured topic in fiction.With the increasing usage and dependence of AI in our society, these concerns are coming closer to reality by the day.AI safety has therefore become a significant field of study, attempting to ensure a safe advancement of powerful and intelligent systems.DeepMind recently proposed a futuristic problem for intelligent AI agents called Reward Tampering, making it possible for agents to bypass their intended objective.This could enable unintended but also potentially harmful behaviours from intelligent and powerful AI systems.This thesis investigates whether Evolutionary Algorithms could be a potential solution or remedy to this problem.The reason why Evolutionary Algorithms might help combat Reward Tampering is that, in contrast to other AI methods, they are able to find multiple different solutions to a problem within a single run, which can aid the search for desirable solutions.Four different Evolutionary Algorithms were deployed in tasks illustrating the problem of Reward Tampering.The algorithms were equipped with varying degrees of human expertise, measuring how human guidance influences performance.The results show a considerable difference in performance between different classes of Evolutionary Algorithms, and indicate that algorithms searching for behavioural diversity are more effective against Reward Tampering and are able to find and preserve safe solutions.However, the results also indicate that the presence of Reward Tampering can limit the performance of algorithms, and therefore inhibit the learning of the desired behaviours.Human expertise also showed to improve the certainty and quality of safe solutions, but was not considered a requirement to avoid tampering.i A special thanks goes to associate professor Kai Olav Ellefsen, for sharing muchneeded insights in the field, valuable contributions, and for being a light for me in dark hours.I would also like to thank my fellow student Mats Tveter, for support, inspiration, laughter, listening to my complaints, and for proving that even the smallest person can change the cause of the future.A big thanks also goes to my family and close friends for endless support and motivation through this work.Also the staff of the Robotics and Intelligent Systems research group for making a great learning environment, and NREC for the computational resources.
Author: Mathias Kvamme Nilsen
Institution: University of Oslo
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7071072338
