When evolving a robot control system to achieve some task, the time taken before a satisfactory controller is obtained is usually dominated by how long it takes to evaluate each variant controller's efficacy at that task (fitness evaluation). The genetic operations take a small amount of time in comparison.
The obvious way to evaluate a control system is to connect it to the robot and see how well it performs in the real world, in real time. Even for tasks that take a short length of time to perform (a minute or two), the large number of fitness evaluations normally required can make this highly time-consuming. Consequently, there is a case for interfacing the evolving hardware control system to a high-speed simulation of the robot and its environment, in order to accelerate the entire evolutionary process.
It is suggested by de Garis[4] that the environment simulation could be implemented in special purpose electronics situated next to the evolving hardware control system on a VLSI chip. The implementation of the simulator in hardware is made feasible by modern automatic synthesis techniques, which can derive a circuit from a textual description resembling a computer program. Implementing the environmental simulator in hardware rather than software makes it faster, but does not solve the problem that it is extremely difficult adequately to simulate the interactions between a control system and its environment, such that a control system evolved in the simulated world behaves in a satisfactory way in the real world. This is especially the case when vision is involved [2]. Nevertheless, it is possible that environment simulation in special purpose hardware will be an important tool as new techniques are developed [13, 23].
When a circuit that has been rapidly evolved for behaviour in a high-speed simulated world is ready for use in the real world, all of its dynamics that influence the robot's behaviour must be slowed down by the same factor by which the real world is slower than the simulation. (Imagine a controller that was evolved for a high-speed simulated world and was then let loose in the real world without being slowed down. Everything in the environment would then be happening slower than it ``expected,'' and the motor signals produced would tend to be too fast for the robot's actuators and the world. It would probably no longer perform the task.) This means that the acceleration of evolution through the use of a high-speed simulated environment is at the cost of the efficiency of the control circuit produced. The final circuit cannot be making maximal use of the available hardware when it is operating in the real world, because it is capable of producing the same behaviour in a world that is running faster: the resources needed to allow for this could be being used for real-world performance. This may, however, be a sensible use of some of the high speed available from electronic hardware.
The fact that it must be possible to adjust the speed of all of the control circuit's dynamics that affect the behaviour of the robot restricts the set of control circuits that could be produced by evolution in a high-speed world. Either the time-scales of all of the semiconductor physics must be adjustable by large amounts (this is not practical), or the aspects of the control circuit's dynamics that make a difference to the robot must be restricted such that they are more easy to control. The latter can be arranged by restricting evolution to use pre-defined indivisible modules that have adjustable time-constants, or by applying the restriction of discrete-time dynamics through the use of an adjustable clock. Each of these possibilities diminishes the degree to which evolution can exploit the resources available from the reconfigurable hardware. Again, this may often be a sensible sacrifice to make for accelerated evolution, but the remainder of this paper deals with an interesting alternative.