1 / 21

AI for Beginners

AI for Beginners. Dr. Thomas R. Ioerger Associate Professor Deptartment of Computer Science Texas A&M University 9/29/2017. AI is not science fiction It is the study of algorithms that we can use to builds systems that produce intelligent behavior There are two broad approaches to AI:

ermin
Télécharger la présentation

AI for Beginners

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. AI for Beginners Dr. Thomas R. Ioerger Associate Professor Deptartment of Computer Science Texas A&M University 9/29/2017

  2. AI is not science fiction • It is the study of algorithms that we can use to builds systems that produce intelligent behavior • There are two broad approaches to AI: • Connectionism - neural networks, deep learning... • a brain is just collection of interconnected neurons anyway • Marvin Minsky, Geoff Hinton... • good for recognition tasks • Symbolic AI - using search and reasoning to produce intelligent behavior • discrete states, concepts represented by symbols, logic-based representation of knowledge, rule-based systems, planning • Symbol Systems Hypothesis (Simon & Newell)

  3. Problem Solving • monkey-and-bananas • farmer-fox-chicken-grain • navigation • stacking blocks

  4. Problem solving as state-space search • define start state, goal, operators (actions) • hard problems = large state spaces

  5. heuristic search • heuristics are functions that guide the search by indicating which states are close to being solved (goal nodes) • A* search algorithm • similar to breadth-first search, but it follows the guidance of heuristic information • the A* algorithm has nice properties like completeness and optimality

  6. Knowledge Representation (KR) • we need ways of encoding knowledge when building intelligent programs • how to represent concepts? • how to reason about actions and outcomes? • Knowledge-based systems (KBS) • aka: rule-based systems • aka: expert systems

  7. KR is typically based on logic • propostional logic • first-order logic • temporal logic • epistemic logic • ... • note: a knowledge base is better than cascaded if-then-else statements • logic sentences are more modular and maintainable and re-purposable

  8. Ontologies • sets of common-sense definitions of... • quantities and volumes; spatial relationships • physical size and shape; fluids; processes • people, organizations • anatomy (SNOMED), diseases (MYCIN) • businesses, industries, finances • media - books, movies, CDs, journals... (Dublic Core XML) • Cyc - a company in Austin started to develop a comprehensive common sense knowledge base "Bill Clinton belongs to the collection of U.S. presidents" • (#$isa #$BillClinton #$UnitedStatesPresident) "All trees are plants". • (#$genls #$Tree-ThePlant #$Plant) • "Paris is the capital of France." • (#$capitalCity #$France #$Paris)

  9. Inference Methods • Satisfiability • search for a truth assignment that satisfies:(A∨B∨¬C) ∧ (B∨D) ∧ (¬A) ∧ (B∨C) ? • Natural deduction • derive new sentences using rules like DeMorgan's Law and Modus Ponens, until generate query • forward-chaining • from known facts, fire rules, draw intermediate consequences, and repeat • backward-chaining • reduce query to sub-goals, until all ground-out in known facts • resolution • proof-by-contradiction; add negation of queyre and convert to CNF; combine to derive new sentences, until derive the empty clause

  10. Example of Natural Deduction Proof showing Colonel West is a criminal 1. x,y,z american(x)  weapon(y)  nation(z)  hostile(z)  sells(x,y,z)  criminal(x) 2. nation(Nono) 3. enemy(Nono,America) 4. x nation(x)  hostile(x,America)  hostile(x) 5. x missle(x)  owns(Nono,x) 6. x missle(x)  weapon(x) 7. x missle(x)  owns(Nono,x)  sells(West,x,Nono) 8. American(West) 9. missle(m1)  owns(Nono,m1) [existential instantiation/skolemization of 5, q={x/m1}] 10. missle(m1) [and elimination on 9] 11. weapon(m1) [modus ponens on 6 an 10, q={x/m1}] 12. hostile(Nono) [modus ponens on 2, 3, and 4, q={x/m1}] 13. sells(West, m1,Nono) [modus ponens on 7 an 9, q={x/m1}] 14. criminal(West) [modus ponens on 1, 8, 2, 12, 11, and 13, q={x/West,y/m1,z/Nono }]

  11. Practical architectures for building KBS • Description Logics (Protege) • JESS • condition-action trigger rules • if carInFrontIsBraking then InitiateBraking • . • Prolog dog(X) :- barks(X). % if something barks, it is a dog has_fur(X) :- dog(X). % dogs have fur barks(snoopy). % a fact ?- has_fur(Z). % a query Z = snoopy % the answer

  12. Planning • We model actions as discrete state transitions • pre-conditions: what must be true in preceding state? • effects: what will be changed in successor state? • example: STRIPS operator for the action of picking up an object X from on top of another object Y • pickup(X,Y): pre-cond: clear(X), on(X,Y), gripper_empty(). effects: holding(X), clear(Y). C

  13. planning: find a sequence of actions that will transform initial state into a goal state • State-Space Search • Goal Regression algorithm • Principle of Means-Ends Analysis - work backwards from goal to find actions that can achieve things not currently true, and the recursively try to achieve their pre-conditions as subgoals

  14. Uncertainty • A critical topic which has been a focus of recent AI research and development is how to deal with Uncertainty • many environments are not fully observable • many environments have stochastic events • many sensors have noise • Methods • probabilistic representations of knowledge • Decision Theory (combine utility + probability of outcomes) • Markov Decision Processes • Bayesian networks

  15. Intelligent Agents • agents are special kinds of AI programs that use some of the above AI methods • common characteristics of agents: • they are situated - they can sense and manipulate an environment dynamically (cause change-of-state) • they are goal-oriented - or at they appear to be trying to achieve goals, e.g. on behalf of human users • they are autonomous - make some decisions on their own • In addition, some agents need to be: • interactive (collaborate in multi-agent systems) • adaptive (learn from experience using machine learning...) • etc...

  16. Ultimately, agents focus on Decision Making • finding actions to achieve goals • this could include inference and planning algorithms • rationality - for each possible percept sequence, a rational agent should select an action that is expected to maximize its performance measure • specialized methods for agents • representing beliefs and intentions of other agents • teamwork - shared goals, roles, responsibilities, trust • collaborative decisions - voting, contracts, auctions... • game theory

  17. Machine Learning a "model" (in this case, a decision tree) that can be used to classify future unseen test examples labeled training data (target concept is: days which people like to play tennis outdoor) A learning algorithm like ID3 PlayTennis=yes

  18. Neural Networks sigmoid threshold unit the parameters in the network are the weights on the edges these must be "tuned" so that they activate the right output nodes given input examples training algorithms (back-propagation): iteratively make small adjustments to weights based on gradient descent until convergence (error stops decreasing)

More Related