We consider a setting where an agent-based plannerinstructs teams of human emergency responders toperform tasks in the real world. Due to uncertaintyin the environment and the inability of the plannerto consider all human preferences and all attributesof the real-world, humans may reject planscomputed by the agent. A na¨?ve solution that replansgiven a rejection is inefficient and does notguarantee the new plan will be acceptable. Hence,we propose a new model re-planning problem usinga Multi-agent Markov Decision Process thatintegrates potential rejections as part of the planningprocess and propose a novel algorithm to efficientlysolve this new model. We empirically evaluateour algorithm and show that it outperformscurrent benchmarks. Our algorithm is also shownto perform better in pilot studies with real humans.
展开▼
