Intelligence evolved in polyworld

In a Google tech video some Alife evolution is demonstrated using the polyworld simulation. This kind of simulated evolution is interesting, but for me the take home message was that really almost no progress has been made in this area in over a decade. I also experimented with similar evolved neural networks in the early 1990s and reached similar conclusions.

The big problem, which still remains something of a mystery and was briefly touched upon during the talk, is that the complexity of the neural network in terms of the variety of behaviors which may be exhibited initially increases rapidly but then reaches a plateau of "no further useful results". In theory it should be possible to evolve very complicated creatures, so why doesn't that happen? One main factor I think is the very low complexity of the simulation environment itself. In real life there are all kinds of different niches which can be exploited in a variety of ways, but in something like polyworld it's all very homogeneous. Biological creatures are also themselves far more complex, being built from millions of cells each of which is a chemical factory with many interactions going on, each of which represents a candidate locus for evolutionary change. When you combine very limited senses and internal dynamics with a very limited environment I think this probably explains why so far we've not seen much of interest going on in the Alife research area recently.

One recent trend in evolutionary algorithms which wasn't talked about is to try to do better than natural selection (or even kin selection) by attempting to characterise the shape of the state space within which the evolutionary search is taking place so that the search can be slightly more intelligently directed. This is a hybrid approach between blind sampled search and more traditional types of search from classical AI.

In other Alife news I notice that Google has blocked Tom Barbalet's Noble Ape site. You can listen to his ragecast here.