Intrinsic Rewards for Maintenance, Approach, Avoidance, and Achievement Goal Types

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Reinforcement learning (RL) relies on task-dependent reward functions, often requiring extensive domain expertise for design.
• Designing effective reward functions is a significant challenge in RL applications.
• Current methods necessitate considerable human input for reward engineering.

Purpose of the Study:

• To propose general, task-independent reward functions for common goal types: maintenance, approach, avoidance, and achievement.
• To introduce metrics for evaluating agent performance in learning these goal types.
• To demonstrate the application of these reward functions in a mobile robot learning framework.

Main Methods:

• Development of general reward functions that leverage inherent goal properties, making them task-agnostic.
• Proposal of performance metrics tailored to each goal type.
• Empirical evaluation within an autonomous goal generation and reinforcement learning framework.
• Testing in a mobile robot application to validate learning capabilities.

Main Results:

• The proposed task-independent reward functions facilitate effective learning across different goal types.
• Empirical results in a mobile robot application demonstrate successful learning guided by the new reward functions.
• The framework successfully generated goals and learned solutions autonomously.
• Compound reward functions were created, enabling learning of complex, sequential behaviors.

Conclusions:

• General, task-independent reward functions can significantly simplify and enhance reinforcement learning.
• The proposed approach reduces the need for domain-specific reward engineering.
• This work paves the way for more adaptable and complex agent behaviors through modular reward design.