Learning Computing History

A Brief History of Artificial Intelligence

Artificial Intelligence (AI) is the science that automates intelligent behaviors.  It is a system that thinks and acts like humans, and rationally.  It is the study of mental faculties through the use of computational methods, the use of computers to do symbolic reasoning, pattern recognition, learning, and some forms of inference.

There are different perspectives of AI:

• Cognitive Science: A way to understand the natural minds and mental phenomena
• Philosophy:  Understand intelligence, and explore some basic philosophical questions as computation
• Engineering: build a intelligent machines

The big question is as Turing asked: “can machines think and behave intelligently?”  Any system that acting humanly requires natural language processing, representation of knowledge, automated reasoning and machine learning.  The Turing test (Loebner contest) rewords $100,000 prize annually for AI system that better than human.

The history of Artificial Intelligence:

1943-1956 Beginnings: Basic principles and generality are general problem solving, theorem proving, formal calculus, and game.  Neural networks among the earliest theories of how reproduce intelligence.

• 1943: Warren McCulloch and Walter Pitts – “ANN” with artificial neurons  and Boolean circuit model of the brain
• 1950: Turing and Shannon chess program; Turing test proposed and debated ever since.
• 1950s Early AI programs, including Samuel’s checkers program, Alan Newell and Herbert Simon’s Logic theorems
• 1951: Marvin Minsky – blocks world, society of mind.  One of AI founder
• 1956:  Dartmouth meeting: “Artificial Intelligence” was coined and adopted, AI established as a discipline.

1952-1969 Early enthusiasm: Focus on search, learning, and knowledge representation.

• 1957: Allen Newell, Herbert Simon – GPS, general problem-solver/planner, early symbolic AI, thought humanly. 
• 1958: John McCarthy’s LISP – time sharing, advice taker had general knowledge and applied new axioms without reprogramming
• 1959: Herbert Gelernter -- geometry theorem prover
• 1960: Widrow & Hoff – perceptron convergence theorem; Terry Winograd – SHRDLU, blocks world.
• 1965: J.A. Robinson invents the resolution principle, basis for automated theorem proving

1966-1974 Bit reality: failed to meet claims of 50’s problems turned out to be hard! Within limited domains, though there are still successful reasoning, perception, and understanding.

• 1965: Weizenbaum’s ELIZA, contained little knowledge of their subject matter, methods worked for simple example, but failed at more complex ones
• Difficulties in automated translation, intractability and limitations of perceptrons discovered, systems for micro-worlds don’t scale up for real application
• 1968: Movie, “A Space Odyssey” brought AI to the public’s attention.
• Micro-worlds success:
  • Evans & Minsky’s ANALOGY could solve IQ test
  • Bobrow & Minsky’s STUDENT could solve algebraic problems
  • Winograd’s SHRDLU could manipulate blocks using robotic arm, and explain itself
  • Minsky & Papert demonstrated the limitation  of neural nets

1969-1979 Knowledge based systems: Knowledge representation schemes and languages became important

• 1969 – 1971, Shakey the mobile robot by Fikes, Hart, Nilsson, a knowledge based systems as opposed to weak method.  Knowledge representation schemes and languages became important
• 1970s: Roger Schank – conceptual dependency theory; Doug Lenat – AM, EURISKO, math discovery, Cyc started to address problems of common sense reasoning and representation; ED Shortliffe, Bruce Buchanan – MYCIN, medical diagnosis, 450 rules, knowledge from experts, better than junior doctors.
• Early expert systems: Ed Feigenbaum – DENDRAL, knowledge intensive molecule structure identification from mass spectrograms.
• Werbos’s and Rumelhard’s back-propagation algorithm held out hope for the ability of AI systems to learn.
• 1979: Hans – Berliner’s heurlstic search player defeated the world backgammon champion.

1980-1988 Used in industry:  Cheaper computing made AI software feasible, AI as the science of intelligent agents, showing relevance to real world applications

• R1 – first successful commercial expert systems, configuring computer systems
• Fuzzy logic and neural networks used in controllers.
• Fifth generation computer system project, logic as the basis for computing, international competition increased research
• Lisp machines and software tools were developed to build expert systems
• Industrial robotic vision systems
• Resurrection  of Neural Nets
• The realization of the system couldn’t do everything

1988-present:  Philosophical extremes in AI, rigorous theorems and experimental rather than intuition, real-world applications

• Building on existing theorems, probabilistic methods – Bayesian nets, Hidden Markov model,
• Base on experimental and strong theories, such has greater reliability, and improved capabilities
• Mathematical formalizations of AI techniques
• Multi-agent systems, internet economies, intelligent agents
• Autonomous systems for space exploration, search and rescue, hazardous environments
• Information retrieval -- Google
• Examples:
  • 1995: RALPH, a CMU’s program drove a van from coast to coast
  • 1997: Deep blue beats the world chess champion
  • 1999: remote agent takes deep space I on a galactic ride
  • 2000: SCIFINANCE synthesizes programs for financial modeling, which develops pricing models, involves the solution of PDEs and more 
  • 2002: Cindy Smart Marketed – a toy can read, tell the time, speech understanding and voice synthesis.
• Key remain areas:
  • Uncertain reasoning
  • Real-time
  • Perception and action
  • Lifelong learning and knowledge acquisition
  • Methodologies of evaluating

Built in 50 years, AI system can do a better job at expert tasks as opposed to mundane tasks that require common sense and experience.  The original question still remain: does a computer think? Why is AI hard? Unfortunately, the last one has bounded by NP – Completeness.