An Introduction to Genetic Algorithms - Boente
An Introduction to Genetic Algorithms - Boente
An Introduction to Genetic Algorithms - Boente
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
function of the average error of the 20 networks over the chosen subset of 20 mappings—low average error<br />
translated <strong>to</strong> high fitness. This fitness was then transformed <strong>to</strong> be a percentage, where a high percentage meant<br />
high fitness.<br />
Using this fitness measure, the GA was run on a population of 40 learning rules, with two−point crossover<br />
and standard mutation. The crossover rate was 0.8 and the mutation rate was 0.01. Typically, over 1000<br />
generations, the fitness of the best learning rules in the population rose from between 40% and 60% in the<br />
initial generation (indicating no significant learning ability) <strong>to</strong> between 80% and 98%, with a mean (over<br />
several runs) of about 92%. The fitness of the delta rule is around 98%, and on one out of a <strong>to</strong>tal of ten runs<br />
the GA discovered this rule. On three of the ten runs, the GA discovered slight variations of this rule with<br />
lower fitness.<br />
These results show that, given a somewhat constrained representation, the GA was able <strong>to</strong> evolve a successful<br />
learning rule for simple single−layer networks. The extent <strong>to</strong> which this method can find learning rules for<br />
more complex networks (including networks with hidden units) remains an open question, but these results<br />
are a first step in that direction. Chalmers suggested that it is unlikely that evolutionary methods will discover<br />
learning methods that are more powerful than back−propagation, but he speculated that the GA might be a<br />
powerful method for discovering learning rules for unsupervised learning paradigms (e.g., reinforcement<br />
learning) or for new classes of network architectures (e.g., recurrent networks).<br />
Chalmers also performed a study of the generality of the evolved learning rules. He tested each of the best<br />
evolved rules on the ten mappings that had not been used in the fitness calculation for that rule (the "test set").<br />
The mean fitness of the best rules on the original mappings was 92%, and Chalmers found that the mean<br />
fitness of these rules on the test set was 91.9%. In short, the evolved rules were quite general.<br />
Chalmers then looked at the question of how diverse the environment has <strong>to</strong> be <strong>to</strong> produce general rules. He<br />
repeated the original experiment, varying the number of mappings in each original environment between 1<br />
and 20. A rule's evolutionary fitness is the fitness obtained by testing a rule on its original environment. A<br />
rule's test fitness is the fitness obtained by testing a rule on ten additional tasks not in the original<br />
environment. Chalmers then measured these two quantities as a function of the number of tasks in the original<br />
environment. The results are shown in figure 2.26. The two curves are the mean evolutionary fitness and the<br />
mean test fitness for rules that were tested in an environment with the given number of tasks. This plot shows<br />
that while the evolutionary fitness stays roughly constant for different numbers of environmental tasks, the<br />
test fitness increases sharply with the number of tasks, leveling off somewhere between 10 and 20 tasks. The<br />
conclusion is that the evolution of a general learning rule requires a diverse environment of tasks. (In this case<br />
of simple single−layer networks, the necessary degree of diversity is fairly small.)<br />
THOUGHT EXERCISES<br />
1.<br />
Chapter 2: <strong>Genetic</strong> <strong>Algorithms</strong> in Problem Solving<br />
Using the function set {AND, OR, NOT} and the terminal set {s –1 , s 0 , s +1 }, construct a parse tree (or<br />
Lisp expression) that encodes the r = 1 majority−rule CA, where s i denotes the state of the<br />
neighborhood site i sites away from the central cell (with indicating distance <strong>to</strong> the left and +<br />
indicating distance <strong>to</strong> the right). AND and OR each take two arguments, and NOT takes one<br />
argument.<br />
60