Figure 1. These two control structures can be "mated" using "chromosomal crossover" as shown in Figure 2 and Figure 3.
Figure 2. The modules of the control structures shown in Figure 1 are arranged sequentially in arbitrary order; this order is part of the genome however and is kept from generation to generation. Chromosomal crossover is simulated by combining these arrays. In this case, the genome is read by starting at the beginning of the chromosome of the first organism. After the yellow module, the reading frame is shifted to the other chromosome by a pseudorandom trigger. The resulting chromosome is shown in Figure 3.
Figure 3. The result of the crossover operation shown in Figure 2. The origins of the communication links ("axons") are kept in place, regardless of any new context.
Figure 4. An example of an efficient control structure consisting of a sweeping sensor controlling a sinus oscillator. The oscillator output is fed back into the offset input, causing exponential decay when the oscillator is switched off. This leads to meandering while inside a food patch, and progressively straighter walking while outside.
Figure 5. Mean fitness as a function of time (yellow curve), and the perceptual number of organisms with food sensors (purple curve) in another experiment. An "adaptive radiation" from generation 0 to generation 10 is followed by a period of "stasis" from generation 10 to 50. Around generation 50, a period of "gradualistic" evolution sets in, where the use of food sensors is fine-tuned. Note that the number of sensor-containing organisms starts to increase long before this makes any visible impact on mean fitness.
Figure 6. Mean fitness (yellow curve) in a run using pure neural networks with only sensors and artificial neurons. Note the punctuation around generation 370, which is linked to the change from edge following to a meandering strategy (Figure 10). The purple curve shows the number of organisms with food sensors, in percent.
Figure 7. Another run using the same parameters as in the run shown in Figure 6, but with another starting point for the pseudorandom number generator. The result is similar, but the punctuation occurred much earlier, demonstrating the instability of the simulations.
Figure 8. The evolution of a strategy involving wide meandering while inside the food patches, switching to a "search" phase while outside. Smaller turning radius is common, but really tight meandering has never emerged in the experiments. All figures in this paper show the entire simulation space of 50 by 50 unit squares.
Figure 9. The evolution of an "edge looping" strategy. Note that just one arbitrary organism from each generation is shown. The organism in generation 5 has been "lucky"; the very tight scribbling would not be efficient with separate, small food patches since the interior of the patch is not utilized. The organism in generation 15 shows a phobotactic behaviour, avoiding its own track, seemingly without being able to sense the presence of food. This behaviour was not seen again later in the simulation, becoming lost in the competition with the edge loopers, Fitness values for the sample organisms shown are 47 (generation 0), 187 (generation 5), 97 (generation 15), 162 (generation 55), 189 (generation 70) and 187 (generation 80).
Figure 10 A-C. Sample trails from generation 320, 390 and 770 in the simulation shown in Figure 6. Note the edge following strategy (but with frequent crossings) before the punctuation around generation 370, replaced by small meandering which is perfected from generation 390 to generation 770.
Figure 10A.
Figure 10B.
Figure 10C.