Towards intrinsic autonomy through evolutionary computation

摘要

This paper presents an embodied open-ended environment driven evolutionary algorithm capable of evolving behaviors of autonomous agents without any explicit description of objectives, evaluation metrics or cooperative dynamics. The main novelty of our technique is obtaining intrinsically motivated autonomy of a virtual robot in continuous activity, by internalizing evolutionary dynamics in order to achieve adaptation of a neural controller, and with no need to frequently restart the agent’s initial conditions as in traditional training methods. Our work is grounded on ideas from the enactive artificial intelligence field and the biological concept of enaction, from which it is argued that what makes a living being “intentional” is the ability to autonomously, adaptively rearrange their microstructure to suit some function in order to maintain its own constitution. We bring an alternative understanding of intrinsic motivation than that traditionally approached by intrinsic motivated reinforcement learning, and so we also make a brief discussion of the relationship between both paradigms and the autonomy of an agent. We show the autonomous development of foraging and navigation behaviors of a virtual robot.