Currently, much of the research on evolved control systems for autonomous agents is centred around simulations of neuro-mimetic networks performed on a general purpose computer. As expertise in the evolution of complex systems advances, so the size and complexity of these networks will increase, with a corresponding decrease in the speed with which they can be simulated in this way. If the speed increase in general purpose computers does not keep pace with improvements in evolutionary techniques, then general purpose computers will have to be abandoned, and more specialised parallel machines used to perform the simulations. Initially, coarse-grain parallel multi-processor computers will be sufficient, but if the complexity of the networks increases still more, progressively finer grained parallelism will have to be utilised, until eventually there is one special purpose processor for each neural node. We would then have arrived at a reconfigurable hardware implementation for artificial neural networks -- evolvable hardware.
The assumption that our ability to evolve control systems will outstrip the speed of simulating computers is not necessarily true. Nevertheless, it is interesting to observe that if our aspirations for building the most complex technically-possible brain-like systems are fulfilled, then the result will be an evolvable hardware implementation of the system. In that case, would it not be possible to evolve hardware in its own right (as an electronic circuit, and not as an implementation of anything else) and obtain an even more powerful system, better suited to silicon than other kinds of architectures like neural networks? I shall return to this question soon.