MotM Submission

Implementing autonomous agent controllers that can robustly and efficiently adapt to different dynamic and complex scenarios is still an open challenge in robotics and AI. For instance, it is a challenge that self-driving vehicles and humanoid robots face because of a continuously changing real-world. Thus, au- tonomous agents have to face multiple sources of stochasticity once deployed. In the last 70 years, many different approaches have been explored to empower a robotic agent with autonomy and adaptivity, aiming to mimic human or animal capabilities.

Inspired by human emotions and their neuro-modulatory effects, some previous work in the literature introduced an emotional module in agent architectures. Precisely, artificial emotions (AE) have been used to modulate the controller’s decisions, and, in turn, the environ- ment guides emotions’ dynamics. In most previous work, the representation of AE has been categorical, enumerating possible emotions and their effects on behavior. Moreover, the given scenarios for previous work have not considered more complex scenarios that require a more specific choice of the rewards. Here we propose the implementation of a RL agent in a survival-based grid world where its goal is to survive as long as possible. The complexity of the given scenario requires the use of deep Q-learning along with Experience Replay as to approximate the Q-values. An architecture is provided that feeds the previously learned policy to a behavioral modulator as to improve the policy by introducing different behaviors.

Our two goals for the work are (1) explore the effects of rewards in a survival game towards understanding the actions taken by the RL agent and (2) achieve effective behavioral modulation by learning the settings in which a chosen behavior is more efficient. The work will build on and extend previous work of the advisor.

• Computer Science

• Computer Science