Hill Climbing with Learning (An Abstraction of Genetic Algorithm)

Abstract Simple modification of standard hill climbing optimization algorithm by taking into account learning features is discussed. Basic concept of this approach is the so-called probability vector, its single entries determine probabilities of appearance of '1' entries in n-bit vectors. This vector is used for the random generation of n-bit vectors that form a neighborhood (specified by the given probability vector). Within the neighborhood a few best solutions (with smallest functional values of a minimized function) are recorded. The feature of learning is introduced here so that the probability vector is updated by a formal analogue of Hebbian learning rule, well-known in the theory of artificial neural networks. The process is repeated until the probability vector entries are close either to zero or to one. The resulting probability vector unambiguously determines an n-bit vector which may be interpreted as an optimal solution of the given optimization task. Resemblance with genetic algorithms is discussed. Effectiveness of the proposed method is illustrated by an example of looking for global minima of a highly multimodal function.

[1]  Anders Krogh,et al.  Introduction to the theory of neural computation , 2019, The advanced book program.

[2]  Zbigniew Michalewicz,et al.  Genetic Algorithms + Data Structures = Evolution Programs , 1992, Artificial Intelligence.