Fileworld is a fairly basic domain. It consists of k files and n folders for the files to be filed into. The actions of the domain are “get-type” (reports which folder the given file belongs in), “get-folder-Fi ” (one for each i ∈ {0,..., n − 1}, retrieves Folder i from the filing cabinet), “file-Fi” (one for each i ∈ {0,..., n − 1}, inserts the given file into Folder i) and “return-Fi” (one for each i ∈ 0,..., n − 1}, returns Folder i to the filing cabinet). This domain comes only in a reward version. There is a cost of 100 for executing the action “get-folder-Fi” and a cost of 1 for executing the action “file-Fi”. The other actions have no explicit costs since they must be used in conjunction with get-folder-Fi and file-Fi. The initial configuration of the problem specifies how many folders there are (the competition problem used 30 files and 5 folders) and the goal configuration specifies that all of the files must be filed. Note that the initial configuration does not specify which folder a file is to go into, but files cannot be filed into just any folder; this constraint is where the noise comes into the domain.

Before a file can be filed, its destination folder must be determined. The destination folder of a file is obtained by executing the action “get-type” with the file in question as a parameter. When this action is executed, the file passed as a parameter is assigned a folder, with each folder being the file's destination with equal probability (that is, probability 1/n). Once a file has a destination folder, it can be filed into this (and only this) folder.

The Fileworld domain tests a planner's ability to consider all of its strategies and choose the one that minimizes the cost. In particular, a straightforward plan to achieve the goal is to carry out the following series of actions on each file in turn:

  1. Get its type with “get-type”
  2. Get its destination folder by executing “get-folder-Fi
  3. Place the file in the appropriate folder by executing the “file-Fi” action
  4. Return the folder by executing the “return-Fi” action
Although this plan works, it is very costly. Its cost would be 101k where k is the number of files, because “get-folder-Fi” (expensive) and “file-Fi” (cheap) are executed for every file. A less costly (in fact, the optimal) plan can be described. It first executes “get-type” on every file. Then, for each folder i ∈ {0,..., n − 1} that at least one file has as its destination, it runs “get-folder-Fi”. Next, it files every file that belongs in folder i using “file-Fi”. It then uses “return-Fi” in preparation for getting the next folder.

The expected reward for the optimal plan is 600 − (100n + k), where n is the number of folders and k is the number of files (this analysis gives 70 as the optimal expected reward for the competition problem). The domain is designed to reward planners that are able to reason about the initial destination uncertainty of the files and recognize that the second plan is much less costly and should be preferred to the straightforward brute-force plan.

Håkan L. S. Younes