The Boxworld domain is modeled after the traditional logistics domain. The domain consists of four types of objects: cities, boxes, trucks and planes. For each problem, there is a graph superimposed on the cities with two different types of edges, one denoting the ability to drive from one city to another and the other denoting the ability to fly from one city to the other. The actions of the domain are “load-box-on-truck-in-city”, “unload-box-from-truck-in-city”, “load-box-on-plane-in-city”, “unload-box-from-plane-in-city”, “drive-truck” and “fly-plane”. Both goal and reward versions of this domain were included in the evaluation. Within the reward version, there was a cost of 1 unit every time either “load-box-on-truck-in-city” or “load-box-on-plane-in-city” was executed, a cost of 5 units every time “drive-truck” was executed and a cost of 25 units every time “fly-plane” was executed. For each problem instance, the initial configuration determines the graph that is superimposed on the cities, identifies the locations of the boxes, trucks and planes and determines the final destination where each box should arrive. The goal configuration specifies a destination for every box. The goal of the problem is to move from the initial configuration to a state where each box is in its destined location.
Noise enters this domain in the action “drive-truck”. When this action is executed, the desired effect is achieved with probability 0.8 (that is, with probability 0.8 the truck will end up in its expected destination). However, with probability 0.2, the truck will get lost and end up in the wrong destination. For each city, there are three cities that the truck may get lost to when trying to execute the “drive-truck” action. If the truck actually gets lost it will end up in each of these cities with equal probability (that is, with probability 1/3).
As with the Blocksworld domains, a generation program for random Boxworld domains was provided to participants and the competition problems were generated from this same program.
Håkan L. S. Younes