Several persons expressed doubts that the network environment by itself, would be adequate to provide selective forces for the evolution of complexity. Manor stated it as the difficulty of competing against small fast replicators like the 22 byte creature. Several suggestions were made as to how features could be added to create selective pressures for evolution toward increased complexity:
Simon suggested using the left-over bits in the soup (in the case of the five bit instruction set, there are three unused bits per byte) to represent some environmental quality, or resource quantity, and to allow the creatures to modify these bits. So the creatures could develop strategies to modify these bits to make the environment where they are located more desireable, and perhaps to attempt to degrade the environment of others. Subsequently, during implementation of these ideas in MacTierra, he found that it was computationally simpler to use a parallel ``environment'' soup, which provides eight bits per byte to play with.
Tarek suggested linking a senescence rate to the real time clock (rather than to the virtual instruction count). This would cause creatures in sleeping soups to degrade an die, further favoring movement.
Manor suggested periodically taking down Tierra, or otherwise killing all ceatures in each soup, in order to select against non-mobile creatures.
Walter suggested (after the workshop) that creatures that receive messages from their offspring on other nodes would be favored, perhaps my manipulating their positions in the reaper queue. This would favor creatures that generate offspring across the net.
Tom suggested that we get some results with evolution in the network environment first, in order to determine if these additional hacks are really needed. Tom favors Manor's suggestion if anything, because it is simple, clean, and would absolutely select for network mobility.
Tom also explained his conception as to how patterns of resource availability on the net could drive evolution to higher complexity. Noise alone probably does not help, since evolution learns patterns, and there is no pattern in noise. But where there are patterns, such as temporal or spatial patterns in the availability of spare CPU cycles, selection could favor behavior that responds by migrating across the net to where and when spare cycles can be expected. However, noise in the form of nodes going down would definetly favor migrational behavior, or at least distributing offspring across the net. Also, the default settings for the network experiment will include a size neutral slicer mechanism, which removes some of the advantage to the small fast replicator that Manor warned about.