Who Makes The Best Partner? Neural Networks With Personalities

Gerald Eisman, San Francisco State University Computer Science Department

Stevenson Hall 1300
12:00 PM - 12:50 PM

Since J.J. Hopfield invented the "energy" function for recurrent neural networks in the early '80s, these systems have been used to find approximate solutions to combinatorial problems such as the Task Assignment problem or the Traveling Salesman problem. Improved results can be obtained by partnering two networks to work on a problem together. The individuals seek their own energy minima but periodically communicate their partial results to one another and then adjust their search accordingly. By giving the networks "personalities" (e.g. stubborn networks refuse to change course, forgetful networks frequently reset and begin again), we find that certain pairs perform better than others.