Transfer learning (TL) is the process by which some aspects of a machine learning model generated on a source task is transferred to a target task, to simplify the learning required to solve the target. TL in Genetic Programming (GP) has not received much attention, since it is normally assumed that an evolved symbolic expression is specifically tailored to a problem’s data and thus cannot be used in other problems. The goal of this work is to present a broad and diverse study of TL in GP, considering a varied set of source and target tasks, and dealing with questions that have received little, or no attention, in previous GP literature. In particular, this work studies the performance of transferred solutions when the source and target tasks are from different domains, and when they do not share a similar input feature space. Additionally, the relationship between the success and failure of transferred solutions is studied, considering different source and target tasks. Finally, the predictability of TL performance is analyzed for the first time in GP literature. GP-based constructive induction of features is used to carry out the study, a wrapper-based approach where GP is used to construct feature transformations and an additional learning algorithm is used to fit the final model. The experimental work presents several notable results and contributions. First, TL is capable of generating solutions that outperform, in many cases, baseline methods in classification and regression tasks. Second, it is shown that some problems are good source problems while others are good targets in a TL system. Third, the transferability of solutions is not necessarily symmetric between two problems. Finally, results show that it is possible to predict the success of TL in some cases, particularly in classification tasks.

中文翻译：

---

遗传编程建设性归纳中的迁移学习

迁移学习 (TL) 是将在源任务上生成的机器学习模型的某些方面转移到目标任务的过程，以简化解决目标所需的学习。遗传编程 (GP) 中的 TL 没有受到太多关注，因为通常假设进化的符号表达式是专门针对问题数据量身定制的，因此不能用于其他问题。这项工作的目标是对 GP 中的 TL 进行广泛而多样的研究，考虑到一系列不同的源和目标任务，并处理在以前的 GP 文献中很少或没有受到关注的问题。特别是，这项工作研究了当源任务和目标任务来自不同域并且它们不共享相似的输入特征空间时迁移解决方案的性能。此外，考虑到不同的源任务和目标任务，研究了迁移解决方案的成功与失败之间的关系。最后，首次在GP文献中分析了TL性能的可预测性。使用基于 GP 的特征构造归纳来进行研究，这是一种基于包装的方法，其中 GP 用于构造特征转换，并使用额外的学习算法来拟合最终模型。实验工作提出了几个显着的结果和贡献。首先，TL 能够生成在许多情况下优于分类和回归任务中的基线方法的解决方案。其次，它表明一些问题是很好的源问题，而另一些问题是 TL 系统中的好目标。第三，解决方案的可转移性在两个问题之间不一定对称。最后，结果表明在某些情况下可以预测 TL 的成功，特别是在分类任务中。