Developmental Learning consists of learning everything from scratch like a newborn baby. At the same time, the agent must discover the structure of its body, learn the structure of its environment, and develop its cognitive capabilities. Implementing an artificial system (robot) capable of such learning constitutes an immense challenge for the field of artificial intelligence.
We will introduce this tutorial by presenting the broad notion of artificial Agent Without Ontological Access to reality (AWOA). An AWOA agent is an agent whose input data is not a representation of the state of reality. Instead, the agent must construct a representation of reality dynamically from regularities observed in its stream of sensorimotor interactions.
Since an AWOA agent has no access to the state of reality or to a predefined representation of reality, it cannot learn to reach predefined states of reality, or to solve problems that are modeled by its designer beforehand. Instead, we expect AWOA agents to engage in open-ended learning by “sedimentation of habits”, which leads to developmental learning.
The game below gives you a glimpse into the problem of constructing a representation of reality from sensorimotor regularities. You are in the situation of a toddler who babbles ignoring the meaning of its sensorimotor experience, learning to gain control of its activity, and constructing knowledge of its world (its own body and its surrounding environment).
If the game does not appear above this line, or for more information, please access it from this page.
This tutorial will teach you how to design agents that can play this game, and more. We will refer to the content presented in the IDEAL MOOC during the fall 2014: http://liris.cnrs.fr/ideal/mooc.