ÄûÃʵ¼º½

Title of the thesis: Towards scalable frameworks for intelligent agents

Thesis defender: Anton Debner
Opponent: Professor Keijo Heljanko, University of Helsinki
Custos: Professor Antti Ylä-Jääski, Aalto University School of Science

Intelligent agents can be utilized in, for example, self driving cars, aerial drones, video games, manufacturing, building automation and communication infrastructure management. Intelligent agents observe their surrounding environment through various sensors and they make intelligent decisions based on these observations. The intelligence of these agents is often based on deep learning techniques that require significant amounts of data to train and test. As training and evaluating intelligent agents in a physical environment can be both expensive and dangerous, it is common to use virtual environments instead. Game engines offer tools for creating such virtual environments, but their scalability has a direct effect on how quickly new agents can be developed and tested. In addition, intelligent agents may require significant computation resources to execute deep learning models. The agents often have to act in real-time, necessitating the computation and communication latencies to be kept within reasonable bounds.

This thesis studies the scalability of game engines and the use of distributed computing for testing and training intelligent agents. The thesis focuses especially on the generation of virtual camera data for autonomous vehicles and the generation of audio data for multiple concurrently acting video game agents that observe their environment through audio sensors. Based on the results, game engines are well suited for training and testing agents, but their scalability could be further extended to better support and speed-up both the development of large-scale networks of multiple collaborative agents and the development of individual agents.

In addition, this thesis examines the computational infrastructure and its configurations related to intelligent agents. We propose a framework consisting of a simulator and a physical computing cluster for researching computing cluster management, quality of service, and energy efficiency in addition to testing the behavior of the intelligent agents. We use a smart city as our example use case for the framework. In the future we expect that the creation of larger systems, such as smart cities, with large numbers of collaboratively acting intelligent agents will put more emphasis on the cost-efficient and rapid development of networks of collaborative agents and the energy efficiency of the related computation for such agents.

Keywords: Intelligent agents, Game engines, Digital twins, Machine learning

Thesis available for public display 7 days prior to the defence at .