• ChatGPT etc. are data-driven AI
    • give it a load of data and let it find the underlying patterns (like machine learning)
    • data + answers = rules
  • we will talk about model-based AI (“the clasiccal AI”)
    • = symbolic AI, rule-based AI
    • programmers define the exact rules and bake them into the system
    • rules + data = outcome

What is an intelligent agent?

  • it has some properties:
    • it has sensors/inputs
      • used to describe the environment around the agent
    • it has effectors (“output”)
      • used to change the environment
    • it performs decision making
      • planning
      • game theory
      • search

Different approaches

  • a matrix introduced in the AI - The modern approach book
  • four different approaches:
    • Thinking humanly - the cognitive science approach. AI models the actual mental processes of humans (how brains work). Closely tied to psychology and neuroscience.
    • Thinking rationally - the “laws of thought” approach. Uses formal logic to represent and reason about the world correctly. Rooted in Aristotle’s syllogisms.
    • Acting humanly - the Turing Test approach. A machine is intelligent if its behavior is indistinguishable from a human’s. Focus on observable output, not the internal process.
    • Acting rationally - the rational agent approach, which the book itself adopts. An agent acts to achieve the best expected outcome given its goals and available information. This is the dominant paradigm in modern AI.

Turing test

  • by Alan Turing, he founded the foundations of Turing machines (and laws of computability), he also helped to resolve the Enigma cypher code
  • A.L.I.C.E - a simple, pattern matching algorithm, that has performed really well of simulating a human behavior
  • this test basically says: “if it acts intelligently, it is intelligent”

A Chinese room argument

  • a program could look smart even though it’s just reading information without understanding them properly (a guy in a room full of Chinese books just reads from them without actually undestanding, what do they mean)
  • it was a philosophical argument - distinguishing syntax and semantics
    • it contradicts the Turing test, if it behaves intelligently, it could just be a sophisticated work with letters and syntax that is far from understanding the semantics of it

History

  • Shakey - a first autonomous robot
    • was able to get the environment around it, plan the task and execute it successfully
  • DARPA Grand Challenge - first competition of autonomous self-driving cars
    • the biggest problem is the explainability of the AI
      • if it is not explainable, we cannot debug it, see, why it decided in the way it has decided
    • the autonomous cars are not that explainable, and that’s a problem
      • the chatbots etc. are also not-explainable
  • Deep Blue vs. Garry Gasparov
  • AlphaGo

Neuro-symbolic AI

  • neuro part: machine learning (gaining knowledge from human experience etc.)
  • symbolic part: the A.I. part: no knowledge, but strong reasoning and explainability
  • the connection of those two parts would be best, but it’s still not developed to some good level
    • it’s an open challenge

Branches of AI

  • it ranges from computer vision, pattern recognition, planning/scheduling, machine learning, robotics, agent systems, text generation, autonomous robots/cars etc.

Tutorial 1

  • add Surynek to Gitlab whenever I want to check my homework and get points
    • if he leaves the project, it means that he has checked it
    • for next assignment, I will have to add him again
  • problem definitions:
    • 2 jugs problem
      • fill the 5l to the full
      • pour 3l to the 3l, leaving 2l in the 5l one
      • empty the 3l one
      • pour the remaining 2l to the 3l one
      • fill the 5l up again
      • and pour the 1l to the 3l one
      • leaving 4 l in the 5l
    • how to move a horse around a chessboard to visit all the cells?
      • for any
    • how to put queens onto a chessboard so they cannot endanger one another?
      • for any
  • state space
    • for these kinds of problems (if we want to transfer them into graph problems):
      • we need to define a state (= vertex)
        • it’s kind of a snapshot of the environment (only the important facts)
      • here it would be an ordered pair (number of liters in first, number of liters in the second)
      • then we need to decide which type of graph it would be
        • (un)directed graph, here it would be directed graph
  • backtracking notes
    • what it is, how is it used etc.