Situated Observation and Participation in Multiple - Agent Systems

A situated agent is not only embedded within its environmental system, but forms an integral and active component of the system as a whole. Accordingly, there are numerous requirements that a situated agent must meet including synchronization with and responsiveness to the system dynamics as well as appropriate and proactive operation — a non-trivial challenge made difficult primarily due to varying environment states and dynamics, difficult tasks with ambitious requirements, and noisy or ambiguous sensory and background information. This thesis addresses these issues by introducing an agent architecture with an interruptible and modular design capable of producing quality-varying solutions based on constraints. The architecture adaptively schedules deliberation processes which enables the agent to evolve its action with the natural frequency of the environment’s dynamics. Resource-bounded deliberation is attained by attributing intentions to the agent, posed in the form of constraints, that influence its action. The thesis demonstrates how these constraints can be designed to achieve complex behaviours. Further, it shows that abstraction of behaviours based on a theory of intentionality provides a comfortable and tractable model not only for behaviour generation but also for behaviour recognition. The theory of intentionality is transcribed into a dynamic, probabilistic model that describes multiple, intentional agents and their environment. It is shown how the properties of intentionality, so transcribed, produce a computationally attractive model that allows real-time approximate inference. Ultimately, a novel agent is introduced for the particular multiple-agent domain of robot soccer. The agent is implemented and tested to study the reactivity of the architecture, accuracy of the environmental modelling, and effectiveness of its deliberation.