An Evolutionary Computation Based Model for Testing Transfer Learning Strategies

To study how Transfer Learning (TL) works and what are effective strategies for transfer learning, we propose to model the TL process using Evolutionary Computation. EC provides a clear model for a problem as searching through a set of potential solutions. We are able to more easily control and measure problem difficulty, problem similarity, and methods of information transfer and relate these to success. As a proof of concept, we will use a static source problem and three fixed target problems with simple known relationships (see Section III). We compare the effectiveness of several ways to transfer knowledge learned from solving one problem to solving the new problems in the context of the relationship between the problems. This we hope will demonstrate that using our EC model is a fruitful way to investigate TL. The results show there is an improvement for using some sampled methods representing the "learned knowledge" of the source problem S. Also, the results show that the diversity of the transferred population has some positive effect on finding the optimal solution depending on the relationship between source and target problems.