A Strike-Search is one way of turning kT-RAM into a stochastic gradient ascent algorithm. It is an interesting computational structure that emerges from the more general Fractal Flow Fabric. The process mimics lighting as it puts ‘feelers’ out into the search space. The evidence accumulates like water from many tributaries in a mountain river. The accumulation starts a positive-feedback process (etches a canyon) that focuses the search into regions with more optimal values. Eventually the process terminates with fixation on a region around a maxima(or minima), the “strike”. Many strikes can be used for a search and strikes can even spawn children, “pre-focused” in regions of the search space.

In the above plot, each red dot is an attempt. As can be seen, the search focuses onto the region of higher value even in the presence of many local maxima. In this case it took 13 attempts before finding 98% of maximum.

HillClimber from Knowm on Vimeo.

 

One thought on “Strike Search (form of stochastic gradient descent)

Comments are closed.