Evaluation Function

• A really important feature of a GA is the Evaluation Function.
• How good is that 20 city gene?
• In general the evaluation function is difficult, but in the TSP case, how would you evaluate a gene?
• You'd just use the length of the path.
• How does the (real) environment evaluate a gene? Survival of the fittest and breeding.
• It is crucial that you build a good evaluation function for your particular problem.
• It's a lot easier if the evaluation function is fast. Why?
• Here's one that's not. I had a student who did a thesis where he wanted to weight rules in a video game agent.
• To evaluate the agent, he took the weights and ran the agent through the game.
• The evaluation function is how well the agent performs.
• What if you have 40 variables to optimize in a neural simulation?
• Here you run the simulation with one set of variables, and see how well it does. This might be faster.
• So, the evaluation function must be task specific.
• Once you've evaluated lots of genes, you make new genes, by combining the best old ones.