Building and Training Autonomous Deep Learning Agents

In the broadest sense, an autonomous agent is anything capable of acting on the environment autonomously. More precisely, it is an entity that operates in the environment without direct action from humans or agents. Most actions of an autonomous agent are sequentially correlated: the current action affects the next action. Many autonomous agents exhibit some form of goal-seeking behaviour; typically these goals are defined through some notion of a reward function or a score function.

Autonomy exists in all forms of life on this planet. People perceive particular events in the environment, often converting the data to an abstract state space. An intelligent person will also choose an appropriate action for the current state. This action may have a long-term effect on a reward score, such as physical health. Of course, people are usually uncertain about their environment and will have a belief, rather than a concrete conclusion, about the best action. Nevertheless, people are capable of choosing their own actions based on their environment.