2.1 Fundamental Issues of Intelligent Systems

(Reading: R&N, chapter 1)

What is Intelligence?

Intelligence:

• “the capacity to learn and solve problems” (Webster dictionary)
• the ability to think and act rationally

Goal in Artificial Intelligence:

• build and understand intelligent systems/agents synergy between philosophy, psychology, and cognitive science computer science and engineering mathematics and physics

What is involved in Intelligence?

• A) Ability to interact with the real world
  • to perceive, understand, and act
  • speech recognition, understanding, and synthesis
  • image understanding (computer vision)
• B) Reasoning and Planning
  • modeling the external world
  • problem solving, planning, and decision making
  • ability to deal with unexpected problems, uncertainty
• C) Learning and Adaptation
  • we are continuously learning and adapting
  • Also: we want systems that adapt to us!
  • Major thrust of industry research.

What is Artificial Intelligence

Rich and Knight: the study of how to make computers do things which, at the moment, people do better.

Handbook of AI: the part of computer science concerned with designing intelligent computer systems, that is, systems that exhibit the characteristics we associate with intelligence in human behavior - understanding language, learning, reasoning, solving problems, etc.

Dean, Allen and Aloimonos: the design and study of the computer programs that behave intelligently.

Russell and Norvig: the study of [rational] agents that exist in an environment and perceive and act.

Riesbeck: AI is the search for answers to the eternal question: AI is a repair process to answer "Why are computers so stupid?" (if most humans can do it, and computers can’t, we have to ask "why?" and "what can we do about that?").

Different Approaches

I Building exact models of human cognition: view from psychology and cognitive science

II The logical thought approach: emphasis on “correct” inference

III Building rational “agents”: agent: something that perceives and acts emphasis on developing methods to match or exceed human performance [in certain domains]. Example: Deep Blue.

Our focus is on III (most recent progress).

Historical Perspective i

Obtaining an understanding of the human mind is one of the final frontiers of modern science.

George Boole, Gottlob Frege, and Alfred Tarski (formalizing the laws of human thought)

Alan Turing, John von Neumann, and Claude Shannon (thinking as computation)

John McCarthy, Marvin Minsky, Herbert Simon, and Allen Newell (the start of the field of AI)

Goals in AI

Engineering Goal : To solve real-world problems. Build systems that exhibit intelligent behavior.

Scientific Goal: To understand what kind of computational mechanisms are needed for modeling intelligent behavior.

Turing Test

Interrogator asks questions of two ?people? who are out of sight and hearing. One is a person; the other is a machine.

30 minutes to ask whatever he or she wants.

Task: to determine, only through the questions and answers typed into a computer terminal, which is which

If can’t reliably distinguish the human from the computer, then the computer is deemed intelligent.

Artificial intelligence is the enterprise of constructing an artifact that can pass the Turing test.

Objections to Turing Test

Newell and Simon [1976]

• Turing test is as much a test of the judge as it is of the machine.
• Promotes the development of artificial con-artists, not artificial intelligence (Loebner competition).

Non-human Intelligence

Why not?

Maybe "intelligence" could be coded any number of ways (biological, mechanical, a collection of wind-powered beer cans, whatever)

Now you might object at this separation of "rationality" from "humanity". You might protest that the only thing that can be rational like a person is another person. And many people would agree with you.

But I don’t think that I am the only kind of thing that can think. That would be like saying that only birds can fly and that air planes, which don’t flap their wings, don’t "really" fly.

What I do think is that there is some abstract notion of flying/thinking that is independent of birds/humans. Like Spock said: "Intelligence does not require bulk, Mr. Scott".

Every computer scientist knows this to be true. Two generations of algorithms research has shown that there exist properties of computation that are independent of what processor the algorithm runs on, or the implementation language. Dijkstra once said "computer science is no more about computers than astronomy is about telescopes"- and he could have been talking about AI.

Not convinced? Well, try another example. Do you think that a robot could/should walk like a human? (see movie of Dinesh Pai’s Platonic Beast). This little fellow walks by occasionally throwing a spare limb over the top of itself. Such a move would tear us apart, but it is the natural way to do it for that kind of walking thing.

And here’s another example:

• Here’s a movie showing two AI algorithms trying to solve the "latin’s square" problem: i.e. pick an initial pattern, then try to fill in the rest of the square with no two colors on the same row or column.
• The deterministic method is the kind of exhaustive theorem proving method used in the 1960s and 1970s that convinced people that "AI can never work".
• The stochastic method is a 1990s-style AI algorithm that makes some initial guess, then refines that guess based on logic feedback.
• This stochastic local search kills the latin squares problem (and, incidently, many other problems).

Now the point of this example is that you would not expect a human to think using stochastic search (too much CPU twiddling). But for a computer, stochastic search is a useful inference method since each local twiddle can be done very quickly.

The Current Frontier

Interesting time for AI

Deep Blue vs. Kasparov (May, ’97)

• first match won against world-champion
• “intelligent & creative” play
• 200 million board positions per second

Kasparov: "I could feel - I could smell - a new kind of intelligence across the table." ... still understood 99.9% of Deep Blue’s moves.

Intriguing issue: How does human cognition deal with the combinatorics of chess?

Different Algorithms, Similar Behavior

Drew McDermott (New York Times, May, 1997): Saying Deep Blue doesn’t really think about chess is like saying an airplane doesn’t really fly because it doesn’t flap its wings. ftp://ftp.cs.yale.edu/pub/mcdermott/papers/deepblue.txt

The brain

• a neuron is the basic processing unit ( 1011)
• many more synapses (1014) connect the neurons
• cycle time: 10-3 seconds (1 millisecond)

How complex can we make computers?

• 108 (Itanium) or more transistors per CPU
• -supercomputer: thousands of CPUs, 1011 bits of RAM
• cycle times: order of 10-9 seconds

Examples, cont

36 Human-Competitive Results by Genetic programming.

Qualitative difference from previous brute-force results.

Does technique generalize? Or is a kludge?

Machine learning

TD Gammon (Tesauro 1993; 1995)

• World champion level but learns from scratch by playing millions of games against itself!
• Has changed human play

ALVINN (Pomerleau 1993)

• Neural net used to steer vehicle in coast-to-coast highway driving
• Speeds of up to 90 mph
• DARPA Grand Challenge / Urban Challenge

Natural Language Processing

BOGOTA, 9 JAN 90 (EFE) - RICARDO ALFONSO CASTELLAR, MAYOR OF ACHI, IN THE NORTHERN DEPARTMENT OF BOLIVAR, WHO WAS KIDNAPPED ON 5 JANUARY, APPARENTLY BY ARMY OF NATIONAL LIBERATION (ELN) GUERRILLAS, WAS FOUND DEAD TODAY, ACCORDING TO AUTHORITIES. CASTELLAR WAS KIDNAPPED ON 5 JANUARY ON THE OUTSKIRTS OF ACHI, ABOUT 850 KM NORTH OF BOGOTA, BY A GROUP OF ARMED MEN, WHO FORCED HIM TO ACCOMPANY THEM TO AN UNDISCLOSED LOCATION.

Summary:

• Date: 05 JAN 90
• Location: COLOMBIA: BOLIVAR (DEPARTMENT): ACHI (TOWN)
• Type: KIDNAPING
• Weapon: *
• Victim: ?RICARDO ALFONSO CASTELLAR? (MAYOR OF ACHI)
• Perpetrator: ?GROUP OF ARMED MEN”
• Organization: ?ARMY OF NATIONAL LIBERATION (ELN)?

Challenges ahead

Note that the examples we discussed so far all involve quite specific tasks.

The systems lack a level of generality and adaptability. They can’t easily (if at all) switch context.

Key issue: knowledge acquisition bottleneck

• Lack of general common sense knowledge.
• CYC project small (Doug Lenat et al.). Attempt to manually encode millions of facts.
• But the Internet is BIG place now. Now possible to automate encoding of facts (machine learning).

The Future

Based on all this, I offer two predictions for the future. One, that we will see a growing number of rational computers but, two, they are going to be aliens (i.e. won’t work exactly like human intelligence) with very different motivations, needs, and desires to us. Instead, the 21st century will see a menagerie of many different kinds of intelligence. Some you’ll know about, like the book-buying assistants wired into Amazon.com that sometimes send you recommendations about what books to read. And some you won’t even see-

• like the Bayesian SPAM filter that eats silly emails,
• like the power-grid specialist that spins up a coal-fired electrical plant since it has guessed that, in two hours time, there will be a spike in the power generation,
• like the NSA data miners rushing SWAT teams to the airport cause it has realized we’re going to be hit with a bomb attack in 20 minutes 19,18,17,....

Think of it as a jungle of AIs, working together, all living in their little ecological niches. And like any ecology, we’ll learn that:

• most can be ignored,
• some have to be to avoided (don’t step on the red ants- they sting)
• while others are useful (if you tie up a flock of swans, you can fly to the moon).

This document was generated on March 1, 2011 using texi2html 5.0.