Neural synthesis of teleo-reactive programs

Abstract

The Teleo-Reactive (TR) formalism has been presented as a new programming paradigm to write agent programs with reactive control and goal oriented behavior. The formalism is based in a circuit semantics that intuitively can be ported directly to a layered neural network architecture. But to capture the essence of the TR paradigm, a mechanism of synthesis must be developed, allowing to express in a neural architecture 1) the reactive nature of the programs. 2) the incremental learning of TR sequences and trees and 3) the continuous feedback from the world. We present an analysis of TR programs and a method to synthesize those programs into an ontogenic neural network model that captures all the features of the program and can evolve with the agent as he explores the world.