Robust execution of belief-desire-intention-based agent programs

Belief-Desire-Intention (BDI) agent systems are a popular approach to building intelligent agents for complex and dynamic domains. In the BDI approach, agents select plans to achieve their goals based on their beliefs. When BDI agents pursue multiple goals in parallel, the interleaving of steps in different plans to achieve goals may result in conflicts, e.g., where the execution of a step in one plan makes the execution of a step in another concurrently executing plan impossible. Conversely, plans may also interact positively with each other, e.g., where the execution of a step in one plan assists the execution of a step in other concurrently executing plans. To avoid negative interactions and exploit positive interactions, an intelligent agent should have the ability to reason about the interactions between its intended plans. We propose SAM, an approach to scheduling the progression of an agent’s intentions (intended plans) based on Monte-Carlo Tree Search and its variant Single-Player Monte-Carlo Tree Search. SAM is capable of selecting plans to achieve an agent’s goals and interleaving the execution steps in these plans in a domain-independent way. In addition, SAM also allows developers to customise how the agent’s goals should be achieved, and schedules the progression of the agent’s intentions in a way that best satisfies the requirements of a particular application. To illustrate the flexibility of SAM, we show how our approach can be configured to prioritise criteria relevant in a range of different scenarios. In each of these scenarios, we evaluate the performance of SAM and compare it with previous approaches to intention progression in both synthetic and real-world domains