| Resumo: | Gene Expression Programming (GEP) is an alternative to Genetic Programming (GP). Given its characteristics compared to GP, we question if GEP should be the standard choice for evolutionary program synthesis, both as base for research and practical application. We raise the question if such a shift could increase the rate of investigation, applicability and the quality of results obtained from evolutionary techniques for code optimization. We present three distinct and unprecedented studies using GEP in an attempt to develop understanding, investigate the potential and forward the branch. Each study has an individual contribution on its own involving GEP. As a whole, the three studies try to investigate di erent aspects that might be critical to answer the questions raised in the previous paragraph. In the rst individual contribution, we investigate GEP's applicability to automatically synthesize sorting algorithms. Performance is compared against GP under similar experimental conditions. GEP is shown to be capable of producing sorting algorithms and outperforms GP in doing so. As a second experiment, we enhanced GEP's evolutionary process with semantic awareness of candidate programs, originating Semantic Gene Expression Programming (SGEP), similarly to how Semantic Genetic Programming (SGP) builds over GP. Geometric semantic concepts are then introduced to SGEP, forming Geometric Semantic Gene Expression Programming (GSGEP). A comparative experiment between GP, GEP, SGP and SGEP is performed using di erent problems and setup combinations. Results were mixed when comparing SGEP and SGP, suggesting performance is signi cantly related to the problem addressed. By out-performing the alternatives in many of the benchmarks, SGEP demonstrates practical potential. The results are analyzed in di erent perspectives, also providing insight on the potential of di erent crossover variations when applied along GP/GEP. GEP' compatibility with innovation developed to work with GP is demonstrated possible without extensive adaptation. Considerations for integration of SGEP are discussed. In the last contribution, a new semantic operator is proposed, SCC, which applies crossover conditionally only when elements are semantically di erent enough, performing mutation otherwise. The strategy attempts to encourage semantic diversity and wider the portion of the semantic-solution space searched. A practical experiment was performed alternating the integration of SCC in the evolutionary process. When using the operator, the quality of obtained solutions alternated between slight improvements and declines. The results don't show a relevant indication of possible advantage from its employment and don't con rm what was expected in the theory. We discuss ways in which further work might investigate this concept and assess if it has practical potential under di erent circumstances. On the other hand, in regards to the basilar questions of this investigation, the process of development and testing of SCC is performed completely on a GEP/SGEP base, suggesting how the latest can be used as the base for future research on evolutionary program synthesis.
|
|---|