Training, Testing, and Generalisation

It is not possible to test a circuit at every point within the operational envelope before assigning it a fitness. Considering the evolutionary algorithm as a machine learning method, during evolution the individuals are evaluated while exposed to a training set of conditions within the envelope, in the hope that the final result will generalise to work over the entire envelope defined for this task. Generalisation must then be tested by checking that the final circuit does work at points within the envelope not experienced during evolution.

A practical method for achieving adequate generalisation is the main goal of the project. For the current initial explorations, the training set consist of five combinations of conditions chosen to represent extremes of a usable operational envelope. Presumably the presence of some extremes in the training set is necessary for adequate generalisation, but it is not yet known if it is sufficient. Analogous difficulties arise in other application domains of evolutionary algorithms; see [3] for an interesting general framework.

It may be that aiming for an adaptive system, rather than a general one, would be more in harmony with an evolutionary approach. For my particular project, this is a choice of viewpoint in interpreting the results, rather than a matter of experimental design. I have defined adequate behaviour at all points within an operational envelope as an engineering requirement, and a selection pressure towards this is provided in the least restrictive way possible. The circuits being evolved have internal state and rich dynamics, so do not necessarily display a constant behaviour over time. This means that although, in response to the selection pressure, evolution could produce a `general' solution in the strict machine-learning sense, it could also produce a circuit which adapts to the current conditions through its own dynamics. If the time taken for this `self-adjustment' were much shorter than the length of a fitness evaluation, then the adaptive circuit would be almost indistinguishable from a general one. Hence evolution is free to explore both avenues; for brevity I will use the term `generalisation' to refer to both possible means of robust observed behaviour throughout this paper.