On The Necessity of Abstraction

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Machine Learning

Background:

• General intelligence necessitates broad problem-solving capabilities, often requiring complex sensorimotor interactions.
• Many tasks demand specific formulations for efficient solutions, posing a challenge for general agents.
• The ability to create task-specific abstract representations is crucial but understudied in achieving general intelligence.

Purpose of the Study:

• To identify and advocate for the importance of learning task-specific abstract representations as a key component of general intelligence.
• To frame the problem of learning abstractions within the reinforcement learning paradigm.
• To review and discuss current advancements in learning action and state abstractions.

Main Methods:

• Conceptual analysis of general intelligence requirements.
• Framing abstraction learning within the reinforcement learning (RL) framework.
• Discussion of existing research on action and state abstraction in RL.

Main Results:

• Task-specific abstract representation learning is identified as a critical, understudied requirement for general intelligence.
• The reinforcement learning paradigm provides a structured approach to studying action and state abstractions.
• Progress and challenges in learning these abstractions within RL are highlighted.

Conclusions:

• Developing agents capable of forming abstract representations is essential for advancing artificial general intelligence.
• Reinforcement learning offers a promising framework for investigating and developing abstraction learning mechanisms.
• Further research into action and state abstraction is vital for creating more capable and flexible intelligent agents.