Surviving the Unexpected - Expect the Unexpected

The short term survival in a stable niche - i.e. against a set of fixed and known competitors etc. is very different from the problem of surviving in the long term where catastrophes and unexpected changes will occur. If these changes are sufficiently severe, so as to restrict
what patterns of strategy can survive, then it is which lineages go extinct that more determines the pattern of evolution than which do slightly better during periods of stability. In other words it is the non-survival of the non-fit that is the determining factor. Here we are distinguishing between a risky situation (i.e. where the probability distribution is know and one is looking to survive improbable events) and an uncertain one (where there are unimagined events) [Knight 1921]. The main question is not how can an individual or group can learn to prepare for the simply improbable but how to prepare for the unimagined.Some simulation models and their results discussed that shed some light on this quesion.

The first of these is a a model of evolving virtual ants living in a uncertain environment. The ants have a rich variety of possible action and genomes and can evolve complex behaviours in a GP manner. The environment is a 3D sandpile where avalanches of unpredictable sizes are possible. There is no exogenous fitness function ants have have eaten enough, be positioned one above the other and both decide to mate for offspring to occur (dont by tree-crossover on all genome). The environment can be tuned so that it has leptokurtic unpredictable avanlanches or be relatively calm. Tentative results weakly support the hypothesis that variety can be selected for and is helpful in such cases.

An artificial stock market, with reactive (dumb) and predictive (smart) traders is examined. The smart traders build models of the stock prices and use these to judge bying and selling conditions, the dumb ones judge future actions on what worked before without trying to predict the future. Results suggest that whilst smart traders spot local trends and gain gradually in that period, they then lose a lot when the market adapts and their model is no longer good. Thus being smart does not give long term advantage here, where the market is always adapting unpredictably to any strategy tried on it.

Edmonds, B. (2002) Surviving on a sand-pile: an investigation as to the type of behaviour that evolves in the presence of crises, Network on Evolvability in Biological and Software Systems Symposium on "Evolvability and Individuality", St. Albans, Hertfordshire, UK. (http://cfpm.org/cpmrep87.html)

Edmonds, B. (2002) Exploring the Value of Prediction in an Artificial Stock Market. Workshop on Adaptive Behavior in Anticipatory Learning Systems 2002, Edinburgh, Scotland, August, 2002. (ABiALs 2002). Butz V.

M., Sigaud, O. and Gérard, P. (eds.) Anticipatory Behavior in Adaptive Learning Systems. Springer, Lecture Notes in Artificial Intelligence, 2684:262-281. (http://cfpm.org/cpmrep94.html).

Frank H. Knight, Risk, Uncertainty and Profit. Houghton-Mifflin 1921