Active Exploration for Learning Symbolic Representations.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Robotics

Background:

• Data-efficient learning is crucial for agents operating in complex environments.
• Current exploration strategies often lack efficiency in model learning.
• Symbolic models offer a powerful abstraction for understanding environment dynamics.

Purpose of the Study:

• To introduce an online active exploration algorithm for data-efficient symbolic model learning.
• To guide agent exploration towards areas of high model uncertainty.
• To evaluate the algorithm's performance against baseline exploration policies.

Main Methods:

• An online active exploration algorithm with two parts: intermediate Bayesian symbolic model generation and uncertainty-guided exploration.
• Utilizing collected data to build an initial symbolic model.
• Directing future exploration based on the model's uncertainty.

Main Results:

• The proposed algorithm significantly outperforms random and greedy exploration policies.
• Demonstrated effectiveness in two distinct computer game domains: an Asteroids-inspired game and the Treasure Game.
• Achieved data-efficient learning of abstract symbolic models.

Conclusions:

• The developed algorithm enables efficient learning of symbolic environment models through active exploration.
• The approach is robust across environments with varying dynamics and logical complexity.
• This work contributes to advancing data-efficient reinforcement learning and model-based AI.