PhD Dissertation Defense of Computer Science Scholar, Asma Sanam

Title: A Framework for Adaptive Decision making for a Team of Autonomous Agents
Friday, May 3, 2019
9:30 a.m.
Room G-13, Aman CED, IBA Main Campus, Karachi, Pakistan
Dr. Sajjad Haider

National Examiners:
Dr. Jawwad A Shamsi, FAST NUCES
Dr. Dr Sharifullah Khan (T.I.), NUST

This dissertation aims to enhance real-time decision making of autonomous agents in a complex adversarial domain. Explicit opponent modeling techniques are applied to store strengths of opponents and to use these strengths to create an opponent model. The devised strategies are optimized specifically for each type of opponent. To deal with changing strategies of the opponents, the strategies are adaptive and revised after predefined time instance. An evolutionary computation-based framework namely SASO has been developed that automates the creation of the opponent model and optimizes strategies specific to each opponent model. The opponent in this research comprises of a team of autonomous robots while the adversarial domain is the simulated soccer platform. For devising opponent-specific strategies, several teams of soccer playing agents have been chosen and their strategies are analyzed. This analysis facilitates in grouping teams into different opponent models to improve gameplay against unseen opponents. The framework proposes a modular approach with a clear distinction between online and offline phases. Both opponent modeling and strategy optimization are performed offline while strategy prediction and strategy adaptation are performed online. Empirical evidence shows that the team, that adapts its strategy according to the opponent, outperforms the team that disregards its opponent.

The challenges addressed in this research are an accurate prediction of the type of opponent, anticipating the opponent’s strategy and then taking correct decisions at the real time. For designing an explicit opponent model, the research proposes past actions of the opponents to build the model. Secondly, there are issues regarding strategy evolution where parameters need to be fine-tuned and a workable strategy has to be guaranteed for all instances. To test the effectiveness of the framework, the RoboCup Soccer Simulation 3D league has been chosen as a test bed. The league offers a dynamic and partially observable environment making strategy recognition and adaptation a truly challenging task.

The novelty of this framework is its end to end approach for strategy extraction, identification, optimization as well as strategy execution in real time to improve the overall performance of the team. It also serves as a generalized approach that prepares agents to interact with unknown opponents. The approach has been implemented over a reasonable number of opponents and can be extended to an exhaustive number of opponent teams.