... (TM)¹

http://www.irobot.com/consumer/roomba_technology.cfm

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... things²

Physical objects. Ssssss. They hurts us, my precious.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... lead³

Only.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... component.⁴

The world simulator might, for example, be used in a MondoSoft game system offering.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... measure.⁵

The enterprising student will be able to show that minimizing trial length is equivalent to maximizing average reward received for goal-seeking agents who receive one unit of reward at the GOAL STATE and zero elsewhere.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... average.⁶

Because the START and GOAL STATEs may be randomly selected, it is only possible to think about the average performance of the AGENT - some TRIALs will inevitably be shorter or longer simply because the START and GOAL are randomly selected to be closer or further. But overall, these fluctuations will average out and the AGENT should be able to reach its goal more quickly.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... non-trivial⁷

Non-trivial here means ``reasonably mixed and vaguely like a real outdoor environment''. Cases like ``one solid band of Grass followed by one solid band of Mud followed by...'' are ruled out. Essentially, if you look at the map of the environment and think ``that could never possibly happen under natural circumstances'', then it's probably a trivial environment.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.