It's all about the 1s and 0s
Anticipation of an opposing team's actions and the ability to generate advantageous responses is an important feature when advising or controlling a team in an adversarial multi-agent system. Such systems include: complex video games, multi-robot systems, and military scenarios. The first step towards doing this is to observe the opposing team and predict their future actions, so that the system has a chance of selecting or advising the appropriate moves for the observing team to perform.
Theories from psychology and neuroscience provide insights into how the human mind performs such predictions, however, there is a large discontinuity between the theories and an actual implementation in a multi-agent AI system. This thesis bridges the gap between a biologically-inspired cognitive architecture based on the simulation theory of mind, and the implementation in a realistic multi-agent synthetic environment.
The basic simulation-theoretic architecture predicts actions by concurrently executing multiple hypotheses and then by finding the hypothesis that best matches the ongoing observed behaviour. This thesis presents work into operationalising this architecture: firstly by applying it to a custom-made real-time-strategy-style game and overcoming the challenges of this domain, which include, for example, the potentially constantly-changing and concurrently-executing goals, and the adaptation of behaviours to both the synthetic environment and the internal models that perform the simulations; secondly by reducing the computational burden of such a wide action-space through filtering and combining potential hypotheses; and finally by defining an attention mechanism to focus observation resources on the most beneficial areas.
The system is shown to be able to predict multiple manoeuvres, formations and targets involving the opponent's agents, from observations in large-scale adversarial environments. Considering the combination of the scalability of the architecture and the techniques for reduced usage of computational resources, it becomes more feasible to use such a principled approach on the current generation of hardware, moving away from ad-hoc approaches that rely on arbitrary measures of the opponent's intentions.