Cellular Automata and Fractal Evolution

Genetic Algorithms and and Circuit Design

Adrian Thompson used genetic algorithms to design a kind of voice-recognition software.
Evolution of electronic circuits became possible with the invention of Field-Programmable Gate Arrays (FPGA), arrays of logic cells whose functions are determined by programs loaded into the FPGA.
Thompson's initial goal was to design a circuit that, given a low-pitch sound (1 kHz) would return no signal, and given a high-pitch sound (10kHz) would return a constant (nonzero) signal.
He began with a population of 50 random circuits and applied the genetic algorithm methodology: assess fitness with a test problem (distinguish low from high pitch), remove the least-fit programs, and let the more fit reproduce by crossover with occasional mutation.
Here is a summary of Thompson's results.
*   After 220 generations, the best circuit could reproduce both sounds.
*   After 650 generations, the best circuit produced a mostly constant signal when presented with a high-pitch sound, but reproduced the low-pitch sound.
*   After 1400 generations, the best circuit produced mostly constant signals in the appropriate ranges.
*   After 2800 generations, the best circuit met the design expectations, and further evolution produced no changes.
A discriminator designed by a person would use a clock to count the time between successive peaks of the input signal.
More clock ticks between successive peaks of the signal means lower frequency and so a 0 should be returned.
Thompson's evolved circuit does not contain a clock.
It appears that the signal is sent through a complicated collection of feedback loops, making a time-delayed reflection of the signal that cancels the low-pitch input, and converts the high-pitch input into a constant. However, there's a surprise.
Some argue if we don't know how a circuit works, we shouldn't trust it with important tasks.
However, we don't know in detail how more traditionally designed microcircuits work, nor are all possibilities tested.
Remember the problem with one of the Pentium chips, only discovered a year after it was marketed.

Return to Genetic Algorithms.