Introduction to AI - Week 11

Dreyfus's Objections

• Explains the failure of existing AI systems to capture expert human judgement.
• Predicts that failure will continue until intelligence ceases to be understood as abstract reason and computers cease to be used as reasoning machines.
• Warns against attempts at too zealous computerisation in fields such as education and management, which, while not AI properly developed, fall prey to similar misconceptions.

Five Stages of Skill Acquisition

• [1.] Novice
  • Recognise various facts and features
  • Novice acquires rules for determining actions.
  • Elements of the situation can be recognised without reference to the overall situation (context-free).
  Example: The beginning automobile driver learning to operate a stick-shift car is told at what speed (context-free feature) to shift gears, at what distance to follow a preceding car.
  Example: Chess rule: "Always exchange your pieces for the opponent's if the total value of pieces captured exceeds that of pieces lost."

Five Stages of Skill Acquisition (Cont'd)

• [2.] Advanced Beginner
  • Experience with real situation
  • Consider many context-free facts.
  • Beginner starts to recognise meaningful elements, which neither an instructor nor the learner can define in terms of objectively recognisable context-free features. Situational elements are considered.
  • Makes use of similarity with prior examples.
  Example: The advanced beginner automobile driver uses situational engine sounds as well as the context-free speed in his gear-shifting rules.
  Example: advanced beginner chess player can spot such situational aspects of positions as a weakened king's side or a strong pawn structure.

Five Stages of Skill Acquisition (Cont'd)

• [3.] Competence
  • The number of recognisable context-free and situational elements present in a real-world circumstance eventually becomes overwhelming.
  Example: competent driver is no longer merely following rules. When he is in a rush, he follows other cars more closely than normally, enters traffic more daringly, and even violates the law.
  Example: competent chess player may decide to attack his opponent's king
  What's missing? "I give instruction to new graduate, very detailed and explicit instructions: When you come in and first see the baby, you take the baby's vital signs and make the physical examination, and you check the I.V. sites, ... When I would say this to them, they would do exactly what I told them to do, no matter what else was going on ... they didn't care if their other kid was screaming its head off."

Five Stages of Skill Acquisition (Cont'd)

• [4.] Proficiency
  • In levels 1-3, decisions are made by conscious choices of goals & decisions after reflecting upon various alternatives.
  • A proficient performer will be experiencing his task from some specific perspective. Certain features are salient.
  • Intuitive ability to use patterns without decomposing them into component features, "holistic similarity recognition"
  Example: proficient driver, approaching a curve on a rainy day, may intuitively realise that he is driving too fast. He then consciously decides whether to apply the brakes, remove his foot from the accelerator, or merely reduce pressure.
  Example: proficient chess player can recognise a very large repertoire of types of positions. Grasps almost immediately, and without conscious effort, the sense of a position.

Five Stages of Skill Acquisition (Cont'd)

• [5.] Expertise
  • An expert generally knows what to do based on mature and practised understanding.
  • An expert's skill has become so much a part of him/her that he/she needs to be no more aware of it than of his/her own body.
  • When things are proceeding normally, experts don't solve problems and don't make decisions; they do what normally works.
  Example: expert driver becomes one with the car, and experiences him/herself simply as driving, rather than as driving a car.
  Example: chess grandmasters can lose entirely the awareness that they are manipulating pieces on a board and see themselves rather as involved participants in a world of opportunities.

Chunks

• Simon has studied the chess master's almost instantaneous understanding of chess positions. They are familiar with thousands of patterns, called chunks. Each chunk is a remembered description of a small group of pieces in a certain relationship to each other.
• Moves spring to mind as chunks are recognised without need for rule-like calculations.

Dreyfus' answer: chess grandmasters recognise and respond to whole positions, not component chunks. (most chess positions consist of several chunks)

Nota bene: 1997 a computer program could beat the world chess champion for the first time (Deep Thought vs Kasparov)

Five Stages of Skill Acquisition

Expert Systems vs. Intuitive Expertise

Roger Shank (1984): The word "expert system" is loaded with a great deal more implied intelligence than is warranted by their actual level of sophistication

Dreyfus: "Actually, we'd prefer to call them "competent systems," since we can find no evidence that they will ever surpass the third stage of our skill model."

Feigenbaum: "The matters that set experts apart from beginners are symbolic, inferential and rooted in experiential knowledge. Human experts have acquired their expertise not only from explicit knowledge found in textbooks and lectures, but also from experience. ..." All the expert system builder need do is extract this knowledge and program them into a computer.

The Chinese Room Argument

• John R. Lucas: Assume a person in a closed room with a rule set for Chinese symbols, he gets Chinese symbols in through a hatch, processes them according to the rules and gives back answers.
  Assumed that the rule set is sufficiently large and comprehensive, the person can simulate competence in Chinese, but he does not understand any Chinese.
• Analogously, Lucas argues, a computer can deal with syntax only, but will never be able to understand the semantics of the notions used.
• But: What about an artificial situated system like a robot?

The Lucas-Penrose Argument

• For any theory rich enough to formulate arithmetic and for any calculus there exists a true theorem which can't be proved by the corresponding calculus. (Gödel's Incompleteness Theorem)
• Since human beings can see the truth of this theorem, human reasoning cannot be based on a fixed calculus.
• Hence human reasoning cannot be mechanised.

But, Do the Arguments not apply to
Machines and Humans alike?

• Gödel: We cannot prove the consistency of human reasoning
• Turing: We cannot assume humans to be consistent

There are problems which are very hard, for machines and humans. We do not yet have a comprehensive understanding of intelligence, but the field is still very young and there is no convincing argument that it is impossible in principle to understand and implement intelligence.

• Zeno's Paradox: Achilles and the Tortoise
  Conclusion: There is no motion.
• Anthony Kenny:
  "John Lucas cannot consistently make this judgement"

Some Problems with Intelligent Software

1997: Boeing changes its general advice for the use of the autopilot in critical situations

What should a machine decide?

• Lauda Air, Boeing B767 crash 26/05/1991: Thrust reverser deployed in flight. The report concluded: "... recovery from the event was uncontrollable for an unexpecting flight crew."
• Lufthansa accident Warsaw 14/09/1993: When is an aeroplane airborne? Should thrust reverser be disabled in all circumstances?

Summary

• AI and Computer Science seem to come closer together. Advances in different areas (e.g., Computer Chess)
• It's difficult to make convincing arguments that something cannot be.
• Extreme care must be taken when AI systems should make decisions of life and death.

Literature    

• John R. Lucas, Minds, Machines and Gödel Philosophy, 1961, p.112-127.

• Hubert L. Dreyfus and Stuart E. Dreyfus, Mind over Machine, 1986.

• Roger Penrose, The Emperor's New Mind, 1989.