Index of Graphs
Overview
These graphs are intended to highlight the differences that I have
discovered while testing the genetic Othello player, factoring
complexity into the fitness measure in one case, and ignoring
complexity when computing fitness in the other. I have created graphs
for all of the categories of the resulting data that I could have;
these seem to highlight the differences between the two approaches the
best. (For example, I have omitted graphs of the depth of the
population because depth is very closely related to complexity, and
I have omitted the complexity of the worst individuals because it is
similar to the data contained in the graphs of the worst fitness and
the average complexity, and the graph itself is rather erratic).
Graphs
Best fitness
This contains the comparison of the fitness measures of the best
individual per five generations between the two methods. Since both methods
produced best individuals with similar fitnesses, the graph is not too
insightful.
Average fitness
This graph contains the comparison of the average fitness measures of
the entire population per five generations between the two methods,
which could be interpreted as the fitness of the entire population at
that generation. This graph reveals quite a bit about the drawback
of constraining the population by complexity.
Worst fitness
This graph contains the comparison of the worst fitness measures of
the entire population per five generations between the two methods.
Like the graph of the best fitness measures, it is not too revealing,
but does demonstrate some possible correlation between complexity and
performance.
Best Complexity
This graph contains the comparison of the complexity of the best
individuals in the populations. While there is not much that can be
drawn from the graph of the best fitnesses themselves, the way that
the population that is evaluated on complexity is constrained to a low
complexity is interesting, while the constrast of the increasing
complexity of the population that was not evaluated on complexity
demonstrates that there may be some benefit (or no loss, at least) in
not evaluating fitness with complexity.
Average Complexity
This graph contains the comparison of the complexity of the average
complexity of the two populations, or how complex the population is at
at a particular generation. When coupled with the display of the
average fitness between the two methods, it makes a compelling case for
the benefits of a population that is not constrained by complexity.
Variety
This graph contains the comparison of the variety of the two populations
over the generations. The interesting point here is the manner in which
the population that is evaluated by complexity is reduced to a small,
repetitive population over many generations. I purposefully made the
complexity factor very small in comparison to the overall fitness measure
(on average, complexity increases the fitness measure by .5 to 1.5
percent, and goes as low as .05 percent for some individuals).
However, over time, complexity in evaluation will grow a very small
population; this, combined with the evaluation of performance, effectively
restricts the search space to a small subset of possible individuals.
Ignoring complexity when evaluating fitness expanded the search space
considerably, leading to a greater variety of individuals in each
generation.
William Bauder
wjb14@columbia.edu