Robotics and Biology Laboratory

Exploring and Exploiting Representations for Intelligent Behavior in 2D Simulation Environments

Motivation

In order to achieve intelligent behavior, agents must effectively gather information about their surroundings and choose appropriate actions. Our hypothesis is agents can efficiently process this information and make robust action selections by identifying and exploiting useful representations relevant to the task at hand. These representations, which abstract relationships between perception and action, may originate from the inherent physical properties of the environment (and the task) or be imposed through feedback control mechanisms. By exploiting regularities in these representation spaces, the evolution of the system state can be constrained, resulting in easier action selection.

The driving insight behind our approach for this thesis is the recognition that not all tasks and environments benefit from a uniform representation of state and action spaces. Some tasks may significantly profit from tailored representations that capture the specific structure and dynamics unique to the problem. In order to gain a deeper understanding and quantify the impact of different representations, we will propose a careful curation of state and action spaces based on specific tasks and analyze the effects of these different representations on reinforcement learning agents.

Throughout this thesis, we will explore various environments and tasks, conducting experiments with diverse representations. Our objective is to conduct an exploratory study aimed at comprehending how distinct representations can influence the performance of agents in simple 2D physics tasks. Additionally, by understanding how varied representations affect the demands of tasks and environments, we aim to produce insights applicable to a broader range of problems.

Description of Work

In this thesis, you will:

  • work on a 2D simulation of an agent in a physical environment, including but not limited to established Gymnasium environments,
  • explore and exploit representations based on the agent, its environment, and the task at hand,
  • identify the optimal next action based on recognized regularities in the representation space.

The scope of work is flexible and not confined to the tasks listed above.

How to apply

You can find all the necessary information here.