Week 4

I. Graph Prototype

• Ported Prof. Rana's Evolution Strategies code to Borland C++.
• Added diversity metric discussed in last week's notes.
• Created primitive graphing tool for ES data. The following three graphs show convergence for several different population sizes.
  • The (mu,lambda) notation is used, where mu is the number of parents, lambda the number of children, and the best mu offspring are chosen as the next parent generation.
  • The maximum (worst), minimum, and average fitness scores are graphed in red, green, and blue, respectively. Diversity is graphed in gray.
  • The x-axis is the number of generations; the y-axis is the fitness or diversity value. Values are mapped onto the y-axis using a Sigmoid scale (the Sigmoid function maps all values from -inf to +inf onto the range 0 to 1). Thus the horizontal lines at the top of each graph are not "cutoffs"; the values are simply so large that their Sigmoid function, when rounded to the nearest pixel, ends up at the top.
  
  ES: (10,100)
  
  
  ES: (10,50)
  
  
  ES: (5,100)
  
  
  
• Obviously it would be useful to have scale information and a legend; I simply haven't gotten to this yet.
• The horizontal lines at the top of each graph are bad; I will have to modify the Sigmoid function somehow to make the mapping even more dramatic.
• It may also be useful to graph the first derivative of these quantities.
• What may work for the scaling function is to have a logarithmic scale (say 10^-3 to 10²⁴), with a special section at the top (10²⁴ to infinity) that is Sigmoid-mapped.