| Summary: | When agents execute a task, their minds focus on the plans, that is, what plans lead to the successful completion of the task, and what plans they are executing with the rest of the team. When there is no guarantee that agents have common knowledge of the task and can observe each other's actions, agents' beliefs about plans can often be incorrect or misaligned, causing them to make uninformed choices of actions that lead to task failure. My research investigates how an agent can be effective within a team, where an agent may be a teammate or a coach.
Previous work has proposed agents that achieve effective coordination by reasoning about a shared flexible plan, assuming common knowledge and full observability of actions, or by reasoning about each other's beliefs about states when those assumptions do not apply. I claim that to be effective in teamwork, agents must reason about the beliefs of their teammates about plans. By recognizing when misconceptions and misalignment in their beliefs about plans might lead to failure and aligning their beliefs as necessary, agents can ensure execution success while minimizing the need for communication.
This thesis provides: (1) A novel modeling framework based on dynamic epistemic logic to represent agents' nested beliefs about plans. This complements existing frameworks that focus on beliefs about states, so that agents can explicitly communicate about plans during coordination. (2) EPike, a computational model for an agent teammate that collaborates with others to execute a task, assuming that agents can observe each other's actions. By planning its actions online in a receding-horizon fashion using an MCTS algorithm, EPike dynamically adapts its actions to its teammates and communicates to align their beliefs. (3) TARS, a computational model for a team coordinator that monitors the team and intervenes when needed to ensure the team's success, when agents may not observe all the actions. By framing the intervention problem as a CC-POMDP, TARS maintains a probabilistic belief about the team's mental state, and only intervenes when the team's risk of failure exceeds a specified threshold. I show the effectiveness of EPike and TARS both in the empirical evaluation and in a VirtualHome simulation testbed.
|
|---|