Definition
In reinforcement learning and artificial intelligence, empowerment is a intrinsic motivation metric that quantifies the amount of control an agent has over its environment. It is defined as the mutual information between the agent’s actions and the resulting future states. By maximizing empowerment, agents are driven to explore environments where they can effect meaningful change, leading to more robust and adaptive behaviors without relying solely on external reward signals.
Summary
A reinforcement learning objective that measures an agent’s ability to influence future states through its actions.
Key Concepts
- Mutual information
- Intrinsic motivation
- Control theory
- Exploration
Use Cases
- Autonomous robot exploration
- Procedural level generation
- Developing general-purpose AI agents