Scientific progress in any field benefits from ``models''
that allow its phenomena to be studied with
a minimum of tedious technicalities and which provide a maximum of
scientific information for the work expended. Genetics has used the
fruit fly Drosophila since about 1910. The motivation for this
is not to breed better fruit flies but is associated with technical
facts like it being possible to store a thousand fruit flies in a
bottle and the generation time being a few days. AI can benefit from
having its own simple ``models''. Here's what looks like a good one
that will be useful for solving many important problems in theoretical
and experimental AI.
The computer game Lemmings (by Psygnosis Ltd.)[Psygnosis 93] provides a framework for formalization of facts about concurrent processes at the level of human interaction with the processes, i.e. when the process with which the human interacts is only partly known and runs too fast for the human to follow. Our examples are all taken from the sampler Lemmings Jr.
The paper concerns using Lemmings as a model for the logical approach to AI including both theoretical and experimental work, but other approaches may also find it challenging and useful. The designers of Lemmings invented games that challenge human ingenuity but keep within a limited lemming physics framework that models an interesting part of the common sense informatic situation. We argue that AI needs just that kind of challenge.
The problems presented by Lemmings are simpler than real-world problems in many important respects. For example, Lemmings is deterministic. The same actions in locally similar situations always have the same local effects. Variations are introduced by limited human ability to repeat the times of actions precisely. However, this is often unimportant.
While much can be learned from theoretical work on formalizing the facts about the Lemmings world, getting the full benefit will probably require actual programs to play the game. In principle, experimenting with programs in a domain where human understanding is adequate should not be required. However, there is so much room for wishful thinking about what facts have to be taken into account that the formalizations are unlikely to be adequate unless they guide an actual program.