L17 Explanation-based Learning.md

• Learn from existing concepts to build new concepts
• Provide novel solutions (creativity)

Concept space

Example: How to prove an object is a cup

• Definition of a cup: an object that is stable and enables drinking
• Description of the object: The object is light and made of porcelain. It has a decoration, a concavity, and a handle. The bottom is flat.
• Concept space for the object:

Prior knowledge

Concepts that the robot already knows:

• A Glass: The glass enables drinking because it carries liquids and is liftable. It is pretty.
• A Brick: The brick is stable because its bottom is flat. A brick is heavy.
• A Briefcase: The briefcase is liftable because it has a handle and is light. It is useful because it is a portable container for papers.
• A Bowl: The bowl carries liquids because it has a concavity. The bowl contains cherry soup.

In the above knowledge representation of a brick, orange arrows indicate causal relationships.

Abstraction

• An AI agent abstracts knowledge from the concepts it already knows
• the knowledge abstracted that are causally related
• The agent is trying to build a causal explanation that will connect the instance (the object bowl) to the cup.
• The agent creates an abstraction of the bowl (so the bowl is replaced by the object)

Transfer

• Example: The agent wants to prove that an object is an instance of a cup.
• The agent construct a causal explanation with causal relationships it abstracted from its prior knowledge.

• In practice, when building the above explanation, the agent goes backward:

  • It starts from the definition of the cup: a cup must be stable and it enables drinking (Planning - [Lesson 13]): two open preconditions that the agent needs to fulfill)
  • Then, it sends a goal of proving the object is stable into its memory. (Problem reduction)
  • Memory returns the brick example because it's stable.
  • It then abstracts from the brick the causal relationship that" "a brick is stable because it has a bottom and the bottom is flat".
  • It applies the above causal relationship to the object to fulfill the first precondition: a cup must be stable
  • Repeat the above steps to build a causal chain to fulfill the second precondition

• What is the minimum knowledge that an AI agent needs to know?

Cognitive connection

• Limitations/considerations in both humans and AI:

  • Humans can only explain things that are consciously accessible (e.g. we can't explain memory processes)
  • Explanation can be post-hoc: Process we use to generate explanation is not necessarily the same as the process we use to arrive at the decision in the first place
  • The process of generating explanation may interfere with the reasoning process

• Generating explanation helps us/ AI agent to learn and understand deeper

• Ability to explain things helps society to accept AI