Adapting the Evaluation Space to Improve Global Learning,
Alan C. Schultz · 1991
In domains where a stochastic process is involved in the evaluation of a candidate solution, multiple evaluations are necessary to obtain a good estimate of the performance of an individual. This work shows that biasing the sampling of that problem configuration space can lead to better performance of the structure being learned given the same amount of effort. 1 Introduction In many domains, particularly those where a stochastic or noisy evaluation process is involved, the evaluation of a candidate solution might require sampling, i.e. multiple evaluations, to get a good estimate of performance. This random sampling over the space of possible configurations of the problem environment is typically performed with a uniform distribution. In some cases, better performance can be achieved by using a nonuniform distribution of samples from this problem configuration space. Furthermore, instead of randomly choosing samples with a fixed distribution, the distribution can be altered adaptive...