Learning Programmatically Structured Representations with Perceptor Gradients

TL;DR

The paper introduces the perceptor gradients algorithm, a novel method for learning symbolic representations from raw data that can be used for planning and control tasks.

Contribution

It proposes a new algorithm that decomposes policies into perceptors and task encodings, enabling the learning of transferable, structured symbolic representations.

Findings

Learns representations compatible with LQR and A* planners

Efficiently learns transferable symbolic representations

Generates new observations based on semantic specifications

Abstract

We present the perceptor gradients algorithm -- a novel approach to learning symbolic representations based on the idea of decomposing an agent's policy into i) a perceptor network extracting symbols from raw observation data and ii) a task encoding program which maps the input symbols to output actions. We show that the proposed algorithm is able to learn representations that can be directly fed into a Linear-Quadratic Regulator (LQR) or a general purpose A* planner. Our experimental results confirm that the perceptor gradients algorithm is able to efficiently learn transferable symbolic representations as well as generate new observations according to a semantically meaningful specification.