Multitask Knowledge Transfer Across Problems

Abstract

Sequential knowledge transfer dealt with the case in which we were focused on tackling a single (target) optimization problem (or task) of interest at a time, by utilizing a static knowledge base of memes learned from various past experiences on related source tasks. Thus, the transfer of memes occurred in a largely unidirectional sense, from whatever had been experienced previously, to the present. In contrast, the case of multitask knowledge transfers caters to distinct tasks of equal priority arising simultaneously. This implies that it may not be possible to await the completion of some tasks for making the acquired knowledge accessible to the others. The different optimization exercises must progress in tandem, with partially evolved memes discovered during the course of the search being spontaneously shared in a dynamic knowledge base. As a result, the transfer of knowledge takes place in all directions (in an omnidirectional sense), with the back-and-forth propagation of memes leading to more synergistic search across multiple problems at once. With this, the main goal of the current chapter is to customize the adaptive memetic transfer optimizer (AMTO), proposed in Chap. 5, for the purpose of multitasking. We label our algorithmic contribution herein as an adaptive memetic multitask optimizer (i.e., AM-MTO).