A history of the Tileworld agent testbed

This paper looks at the history and development of an agent testbed called Tileworld. It defines the original testbed and documents its gradual development up to its current form. It also introduces some of the experiments performed using Tileworld and their results. It concludes with a comparison of Tileworld with other agent testbeds and a look at agent testbeds as a whole. Tileworld A History Tileworld was initially introduced in [Pollack and Ringuette, 1990] as a system with a highly parameterised environment which could be used to investigate reasoning in agents. The original Tileworld consists of a grid of cells (squares) on which various objects can exist. These objects can be anyone of the following, agents, tiles, obstacles and holes. The agent (about which the experiment is based) can move up, down left or right. The agents objective is to pick up and move tiles so as to fill the holes. A hole has an associated point value which is awarded to the agent upon filling the hole. Each hole varies in size and point value, so a hole may consist of three cells on the grid and have a total point value of five. Once the hole is filled completely the agent gains the points. The agent knows how valuable each hole is in advance; its overall goal is to get as many points as possible.Tileworld simulations are dynamic, the environment changes continually over time. The objects appear and disappear at rates pre-determined by parameters in the simulator. The first use of Tileworld as a testbed was in [Pollack and Ringuette, 1990] in which it was used to test the IRMS architecture [Bratman et al., 1988]. Following the initial experiments in [Pollack and Ringuette, 1990] tileworld was shown to be a viable testbed for evaluating agent architectures. In [Kinny, 1990] Kinny investigates various simulated worlds and sets out four properties an idealised simulated world should have. • A set of objects and events sufficiently rich to embody ”interesting” aspects of real environments