Generating Natural Behaviors using Constructivist Algorithms
Proceedings of the First International Workshop on Self-Supervised Learning, PMLR 131:5-14, 2020.
We present a project to design interactive devices (smart displays, robots, etc.) capable of self-motivated learning through non-goal-directed interactive behaviors (e.g., curious, emotional, playful behaviors). We use and improve algorithms inspired by constructivist epistemology that we have designed previously. These algorithms incrementally learn se- quential hierarchies of control loops in a bottom-up and open-ended fashion, and continu- ously reuse the learned higher-level control loops to generate increasingly complex behaviors that exhibit self-motivation. This project contributes to research in self-supervised learning because the learning is driven by low-level preferences that under-determine the device’s fu- ture behaviors, leaving room for individuation, which, in turn, opens the way to autonomy in learning.