CCM, genetic algorithms (GA), neural networks such as Hopfield's [Hop 85] and constraint programming all target algorithm-less computation. CCM and GA are similar in that both perform stochastic computation, and that data are converted and tested by evaluation functions. However, they differ in that the evaluation functions in GA use global information, but those in CCM use local information only. GA is applicable only when a specification can be written. Hopfield networks also work with global evaluation functions called energy functions, though they are not computed explicitly. In constraint programming, symbolic constraints are usually satisfied by constraint propagation. The N queens system in CCM suggests an alternative method of constraint satisfaction, which is much easier to program and probably more robust.
CCM-based systems that use a parallel scheduling strategy have similarity to the chemical abstract machine [Ber 90] which is also based on a production system. However, their target is not self-organization but the building of a semantic model for parallel computation, and it does not use evaluation functions. A set of rules and LODs in CCM can also be regarded as a description of a probabilistic algorithm [Bra 88]. However, our target is quite different from that of conventional probabilistic algorithms. A probabilistic algorithm must have correctness, which is meaningful only when there is a specification of the algorithm. On the contrary, our target is computation without complete specification.
Other approaches related to the self-organization of computational systems include self-organizing neural networks, artificial life, and connectics [Tak 91].