- Keywords: Informed Search, Deep Learning, Heuristics, Transfer Learning, Efference Theory
- TL;DR: We present a neuroscience-inspired method based on neural networks for latent space search
- Abstract: Planning in high-dimensional space remains a challenging problem, even with recent advances in algorithms and computational power. We are inspired by efference copy and sensory reafference theory from neuroscience. Our aim is to allow agents to form mental models of their environments for planning. The cerebellum is emulated with a two-stream, fully connected, predictor network. The network receives as inputs the efference as well as the features of the current state. Building on insights gained from knowledge distillation methods, we choose as our features the outputs of a pre-trained network, yielding a compressed representation of the current state. The representation is chosen such that it allows for fast search using classical graph search algorithms. We display the effectiveness of our approach on a viewpoint-matching task using a modified best-first search algorithm.