David J. Musliner
In the beginning, there was always a plan. Monkeys and bananas, cannibals and missionaries, blocks world.., the planner just had to be smart enough, and look hard enough, and there was a plan to be found. Even in "real world" problems, there was always a plan: you can always find a way to get from Princeton to Brown, even if the airport is closed. If the planner couldn’t find a plan, either the planner was broken or the domain description was broken. Building complex domain descriptions was still hard, because planners would say little about why they couldn’t build a plan. But now the problem is much harder. Now, planners are going to control spacecraft and refineries and autonomous combat aircraft, and there may not always be a plan. Parts break, sensors fail, accidents happen, adversaries thwart intentions. The planner (or at least the overall agent control system) will have to make tradeoffs: if the goal to take high-resolution pictures cannot be satisfied, take low-resolution pictures; if the pump breaks and temperature skyrockets, reduce throughput; if the enemy may shoot you down, fly as carefully as you can.