Artificial Intelligence Terminology

From sheep
Jump to navigation Jump to search

Terminology

Intelligent Agent
Interacts with environment

Sensors -> Control Policy -> Actuators = Perception Action Cycle

Fully Observable
Sufficient information to make optimal decision
Partially Observable
Memory required on agent - not all information available
Deterministic
Agents actions uniquely determine outcome
Stochastic
Outcomes include randomness - not wholly predicted by action e.g. dice game
Discrete
Finite number of sense/actions e.g. chess
Continuous - Infinite number of sense/actions e.g. darts
Benign
No others have objective that is contra to your objective e.g. weather
Adversal
They are out to get you - e.g. Chess
Activity Fully Observable Stocastic Discrete Adversal
Chess Y N (Determined by Agent + Opponent) Y Y
Poker N Y Y Y
Robotic car N Y N N (Not in Stanford)

Rationality

Does the "right" thing given what it knows

Uncertainty management in software

What to do when you don't "know"

  • Causes - Sensor limits, Adversary, Laziness (could compute - e.g. too expensive), Ignorant - could know but don't care to know

Problem Solving

  1. Initial state
  2. Action(State) -> {a1, a2, a3...} - actions from set may be dependent on state
  3. Result(State, Action) -> S' - new state
  4. GoalTest(State) -> True/False - reached goal
  5. PathCost(S -a1-> S' -a2-> S''...) = n - cost of step, action combinations
  6. StepCost(S, a, S') = n - Most path costs can be considered additive thus this can be considered the sum of a StepCost

Admissible heuristics

heuristic must be optimistic

  • h(s) < true cost

Problem solving requirements

  • fully observable
  • known domain - actions available
  • discrete domain - finite
  • deterministic
  • static - nothing else changes world other than our actions

Turing Test

Requires:

  • natural language processing
  • knowledge representation
  • automated reasoning
  • machine learning

Full Turing test also:

  • computer vision
  • robotics

Syllogisms

  • logical reasoned arguments

Logic solving systems exist but modelling a problem 100% logic and resources required are limiting factors in using this