Imagination boosts strategy learning in artificial and human agents

Despite varied studies of imagination in natural and artificial learning systems, we don't yet have an adequate understanding of the overarching computational utility of imagination. In Chapter 1 of this work I offer a structural definition of imagination based on prior literature. I then characterize imagination’s relationship with the closely related process of memory replay, introducing the concept of the “replay-imagination continuum”. From these characterizations follow my hypothesis that, in relation to memory replay, imagination should be most useful for learning and strategy generalization in environments that are only partially-explored. I test this hypothesis by comparing imagination-based versus replay-based training on strategy learning in impartial combinatorial games using artificial agent simulations (Chapter 2, Aim I) and human behavioral experiments (Chapter 3, Aim II). In chapter 4 I synthesize the overlapping results from the artificial agent and human experiments. In both agent types, imagination boosts performance in relation to memory replay, indicating that imagination aids strategy learning as a general computational utility.