17 - Emergence of tool construction and tool use through hierarchical reinforcement learning

Abstract

Tool use and tool construction are important indicators of intelligence and high-level cognition. Various animals have been found to use rudimentary tools, and in limited cases tool construction has been observed. However, humans are by far the most advanced in terms of tool construction and use. Considering the potential impact tools have had on the enhancement of human intelligence and vice versa, understanding the nature of this kind of brain-tool interaction can provide us with key insights on how to build self-improving AI. In this chapter, as a first step, we investigated primitive tool construction and use using deep hierarchical reinforcement learning. The results show that with minimal reward shaping, an agent can learn to construct and use a simple tool. These are primitive results, but we hope they can serve as a steppingstone toward a full-blown tool-intelligence coevolution for open-ended AI.