Learning to Predict Consequences as a Method of Knowledge Transfer in Reinforcement Learning

Area of Science:

• Artificial Intelligence
• Machine Learning
• Robotics

Background:

• Reinforcement learning (RL) agents learn through trial-and-error.
• Efficient long-term learning requires transferring knowledge to new tasks.
• Predicting action consequences is key to effective knowledge transfer.

Purpose of the Study:

• To propose a novel knowledge transfer method for RL agents.
• To leverage agent-centric information for predicting environmental consequences.
• To improve learning efficiency and reduce costs in novel environments.

Main Methods:

• Developed an RL approach using both agent-centric and environment-centric information.
• Trained agents to predict action consequences based on agent-centric data.
• Evaluated the method on spatial navigation and network routing tasks.

Main Results:

• The proposed knowledge transfer approach demonstrated faster learning.
• The method resulted in lower learning costs compared to alternatives.
• Agent-centric predictions effectively transferred knowledge to environment-centric learning.

Conclusions:

• Predicting action consequences using agent-centric information facilitates efficient knowledge transfer in RL.
• This approach enhances an agent's ability to adapt to new tasks and environments.
• The method shows significant advantages for complex domains like navigation and routing.