Semantic Similarity Based Crossover in GP: The Case for Real-Valued Function Regression

Semantic-Based Subtree Crossover Applied to Dynamic Problems

“…SSC has been extended to the Most Semantic Similarity based Crossover (MSSC) [26] to predict time series and the experimental results in [26] show that MSSC helps to improve the performance of GP in solving this problem. However, the problems in the previous studies [27], [26] are static problems. Therefore, it raises a question whether these semantic-based crossovers still maintain their superior performance when they are applied to dynamic problems?…”

“…In the field of Genetic Programming (GP) [24], [20], researchers have recently paid more attention to semantic information, with a dramatic increase in the number of publications (e.g., [14], [16], [17], [19], [18], [1], [23], [25], [27], [8]). Previously, most research has purely focused on syntactic aspects of GP representation.…”

“…In a recent work, Uy et al proposed a Semantic Similarity based Crossover (SSC) with the aim to improve semantic locality [27]. The basic idea of SSC is keeping a semantically small change in crossover by exchanging two semantically similar subtrees.…”

“…The basic idea of SSC is keeping a semantically small change in crossover by exchanging two semantically similar subtrees. The experiments in [27] show that SSC helps to significantly improve GP performance in comparison with standard crossover in solving a class of real valued symbolic regression problems. SSC has been extended to the Most Semantic Similarity based Crossover (MSSC) [26] to predict time series and the experimental results in [26] show that MSSC helps to improve the performance of GP in solving this problem.…”

Searching for novel regression functions

“…Such problems have been extensively studied and analyzed by the GP community [22]. Recently, symbolic regression GP systems have been improved by incorporating the concept of semantics [23], [24], [25], [26], [27], [28], [29], [30], [31]. On the other hand, the present work takes an approach based on behaviors, the difference between both approaches is carefully discussed below.…”