Week 19

• The following link will take you to a draft version of the "Program Design" chapter of my report.

• I have implemented the weights-as-sliders feature in the network visualization window. You can slide the weights up and down, and watch the fitness change as you do so. This is interesting in itself, but what is even more interesting is that, if you slide the weights while the backpropagation algorithm is running, you can watch the effect that individual weights have on the gradient-based search. For example, sometimes when you slide a weight while backpropagation is running, the weight will "snap back" to its original position, as the backpropagation algorithm returns to the same local optimum it was in before the weight changed. But other times, changing a single weight will cause many other weights to change in response. If you think of the local optimum as a "hill", then fiddling with the various weights can give you an estimate of how far the hill extends along each of the "axes" in weight space.
• I encourage you to try this out for yourself. The following link will allow you to download an updated demo of my program. One caveat: right now, the fitness score is displayed on the title bar as it changes; I will soon change this to a "Fitness: . . . " text field within the window.