"FITNESS SCALING" IS WRONG .... MAYBE.
- Speaker: Dr Tom Westerdale
- Date: Wednesday, 23 May 2012 from 16:45 to 17:45
- Location: Room 745, Malet Street
We address a question that arises in many evolutionary systems, but is clearest in genetic algorithms. An individual's VALUE is a measure of its usefulness. Reproductive rate is a function of value, but what should that function be? John Holland and others have long tried to show that reproductive rate should be proportional to value, but I find their arguments unconvincing. In the talk, I shall present my argument that proportionality minimizes dangerous sampling error noise. (Such noise can lead to premature convergence.) "Fitness" sometimes means value and sometimes means reproductive rate. "Scaled fitness" is genetic algorithms jargon for reproductive rate that is not proportional to value, so my argment implies that "fitness scaling" is dangerous.
My argument is in two parts, both of which are necessary to establish any of the result. I will present the asexual part first, to be sure we have time for its rather nice formal argument, which is simple and rather pictorial. Then we will discuss the sexual part, which I will have time only to outline. I will indicate the weakest point in this part. Time permitting, we will then discuss related topics, such as: (A) Significance and limitations of the result. (Is "fitness scaling" always wrong? No.) (B) Why biologists missed this argument, or if they saw it, got it upside down. These related topics come last because they are open ended.
The argument is unpublished and unrefereed, so I would appreciate your help in spotting errors.