Week 15

I. A Busy Week

• Between finals, final projects, and coming down with the flu over the weekend, I did not have much time to work on my MQP this week. One thing I did do was to add generation numbers next to the matrix displays in the "Generations" window:
  
  
  
  This makes it easier to keep track of what you're looking at, especially when zooming in or out.

II. Term Summary

• Perhaps my most important accomplishment this term was making my program dramatically more interactive. I have added the ability to browse generations, to view the family tree for any network, to save key networks for later use, and to adjust the level of mutation while the evolution is in progress.
• I have made my network display more informative by using a combination of line length and brightness to indicate weight strength, and have added perpendicular lines to indicate the volatility of each weight. In addition, I came up with a compact matrix display which allows many networks to be viewed and compared at once.
• Finally, I have solved at least two non-trivial problems, namely the scale-balancing problem and the problem with housing data described in the previous week's notes.

III. Observed Utility

• I have found my network display to be useful in spotlighting unnecessary hidden nodes. When all the weights going into or out of a given node are near zero, I have been able to optimize the search process by reducing the number of hidden nodes.
• When I was trying to design neural networks by hand, I found the Network Editor to be a useful debugging tool. By selecting specific hidden nodes and monitoring their inputs and outputs for different samples, I was able to see which instances in the training set were causing problems.
• By looking at the "Generations" window and examining the family trees for different networks, I was able to better understand the phenomenon of genetic drift.

IV. Future Goals

• In the coming term, I would like to further increase the interactivity of my program. I think it would be useful to be able to treat the weight and sigma bars as sliders, and to be able to slide a weight back and forth and watch the fitness change on the fly.
• You should be able to apply back-propagation to individual networks or to the entire population.
• It would be nice to be able to load & save entire sessions, rather than just individual networks.
• Finally, I would like to solve more problems from the Machine Learning Repository.

V. Executable Demo (Again)