History of AI: 1934 to Present

AI History (1934-Now):

• From 1934-1955, AI's gestation period included Warren McCulloch and Walter Pitts's first artificial neuron model and Hebbian learning as groundwork for neural networks.
• It was theorized that any computable function could be modeled by a set of neurons.
• In 1950, Turing introduced the Turing test and concepts of reinforcement and machine learning in "Computing Machinery and Intelligence."
• In 1956, the Dartmouth meeting marked AI inception.
• The "Logic Theorist" AI program proved theorems.
• AI showed capabilities in games, puzzles, and IQ tests from 1952-1969, rivaling human uniqueness.
• The General Problem Solver (GPS) mimicked human problem-solving.
• Arthur Samuel’s checkers program demonstrated early machine learning successes.
• The AI programming language LISP was defined.
• The resolution method (logical reasoning) was discovered in 1965 by Robinson.
• The AI Winter (1966-1973) included computational intractability and Minsky and Papert’s "Perceptrons" exposing limitations of simple neural networks.
• Neural networks began to disappear.
• From 1969-1979, knowledge-based systems used domain knowledge to allow for stronger reasoning.
• AI became an industry starting in 1980.
• The Digital Equipment Corporation sold the R1 "expert system."
• Rule-based expert systems like DENDRAL and MYCIN rose in prominence.
• The industry grew from a few million to billions in 8 years.
• From 1986 onward, neural networks returned with the back-propagation algorithm.
• AI adopted the scientific method around 1987.
• Speech recognition and HMM became common.
• Intelligent agents emerged around 1995, including search engines and recommender systems.
• Large data sets became available around 2001.
• Deep learning revolutionized AI in 2011, excelling in speech and image recognition.
• Its 2012 ImageNet contest victory expanded AI to various domains.
• This resurgence, fueled by advanced hardware and data, led to widespread AI optimism.

State of the Art:

• Achievements include robotics vehicles in the DARPA Challenge and advances in speech recognition.
• Autonomous planning and scheduling examples include Remote Agent for spacecraft and MAPGEN for NASA's Mars rovers.
• AI is used in game playing and spam fighting.
• Logistic planning tools include DART.
• Robotics has made major advancements.
• Machine translation uses statistical models.
• Generative AI includes Large Language Models like ChatGPT and Gemini, as well as image/video generation.
• Waymo vehicles reached 10 million miles, autonomously, without serious accidents.
• Legged locomotion advancements include robots like BigDog and Atlas.
• NASA's Remote Agent is used for spacecraft and EUROPA toolkit for Mars rover ops.
• Online systems translate documents into 100+ languages.
• Microsoft's speech recognition has achieved a 5.1% word error rate.
• Personalized recommendations are powered by AI on platforms like Amazon, Netflix, and Spotify.
• AI surpassed human players in games like Go and Chess.
• AI excels in object recognition and image captioning.
• AI matches/exceeds expert doctors in image-based disease diagnosis, enhancing medical diagnostics and treatments.
• Deep learning uncovers detailed information on extreme weather events, contributing to climate change research.
• AI is used as a tool to advance and transform science, for example, within biochemistry.
• AlphaFold helps determine protein structures.
• The CASP Competition involves DeepMind.

Intelligent Agents:

• This refers to a framework for designing AI systems.
• The four basic types include simple reflex, reflex with state, goal-based, and utility-based agents.

Agent Types:

• Simple reflex agents respond directly to percepts.
• Reflex agents with state maintain an internal model to track aspects of the world.
• The agent maintains an internal state, updating its knowledge.
• Condition-action rules determine the best response based on the current state.
• Effectors take action, influencing the environment.
• Goal-based agents act to achieve their goals using a model of the world and a set of goal states.
• Utility-based agents choose actions based on maximizing expected utility using a utility function.

Properties of Task Environments:

• Environments can be fully or partially observable.
• Environments can be single-agent vs. multi-agent (competitive or cooperative).
• Environments can be deterministic vs. stochastic, and episodic vs. sequential.
• Environments can be static vs. dynamic, discrete vs. continuous, and known vs. unknown.
• The "hardest case" is an environment comprised of partially observable, multiagent, stochastic, sequential, dynamic, continuous, and unknown factors.

Key Concepts for Agents:

• Agents perceive and act in an environment.
• An agent function specifies an agent's action in response to a percept sequence.
• The performance measure evaluates agent behavior.
• A rational agent maximizes expected performance given its percept sequence.
• Agents can improve their performance through learning.
• A task environment specification includes the performance measure, environment, actuators, and sensors.
• Task environments vary and can be fully/partially observable, single/multiagent, deterministic/nondeterministic, episodic/sequential, static/dynamic, discrete/continuous, and known/unknown.
• The agent design should reflect uncertainty about objectives when performance measures are unknown.
• The agent program implements the agent function, with designs varying in efficiency and flexibility based on the environment.

Work Schedule topics include:

• CMPG chatbot comparison.
• Computational resources for Intelligence.
• Teachable machine.
• Lobe.ai.