Post on 19-Dec-2015
CS498-EACS498-EAReasoning in AIReasoning in AI
Lecture #1Lecture #1
Instructor: Eyal AmirInstructor: Eyal Amir
Fall Semester 2011Fall Semester 2011
Artificial Intelligence (AI)
• Subfield of computer science• Three related tasks
– Understand “intelligence”• Cognitive theories, architectures• Similar fields: linguistics, cognitive science
– Build “intelligence”• Commonsense reasoning
– Apply “intelligent” techniques to applications• NLP, Econometrics, Social Networks, Robotics,
Electronic commerce
Artificial Intelligence (AI)
Reasoning
NaturalLanguage
Learning
Vision
Knowledge
DecisionMaking
Robotics
Artificial Intelligence (AI)
Reasoning
NaturalLanguage
Learning
Vision
Knowledge
DecisionMaking
Robotics
Reasoning in AI
• Reason with a given knowledge• Types of knowledge:
– Logical: (p & q r)– Probabilistic (Pr(p,q) = 0.75)– Preferences (p preferred over q)– Data and Observations: “p occurred”
• Knowledge structures– Space, time, actions, situations– Beliefs, knowledge (of agents)– State representations
Reasoning in AI
• Reason with a given knowledge• Types of reasoning:
– Does r follow from p? What can explain p?– What is the probability of r? What value of r is
most likely? Mean, Mode, Median of r?– Do I prefer p over q? What if r holds?– Did action a occur? What sequence of events
is most likely?– Is a model correct? Likely?
Reasoning in AI
• Decision Making and Planning are special cases of Reasoning in AI
• We will cover them only very briefly– They are the topic of other courses– They can be cast within the frameworks that we will
discuss here
• Decision Making as Reasoning– Given action descriptions (knowledge)– Find (reason) a sequence/policy of actions that
achieves the goal or optimizes our rewards/value
Reasoning in AI
• Learning can also be seen as a special cases of Reasoning in AI
• Bayesian Learning = reasoning about parameters in a model that includes observations– Estimating the parameters of a model =
learning– Finding the most-likely parameters or the
expected parameters = reasoning
AI Applications
Reasoning
NaturalLanguage
Learning
Vision
Knowledge
DecisionMaking
Robotics
Medicin
Econometrics
SocialScience
Databases
Networks
AutonomousVehicles
ElectronicCommerce
Econometrics
• Capital Asset Pricing Model (CAPM):– E(Ri) = Rf + bi*(E(Rm)-Rf)– bi – the sensitivity of the asset returns to the
market returns– E(Rm)-Rf – risk premium
• Estimating bi from data – learning [reasoning]
• Using regression in a linear model:– Bi = Cov(Ri,Rm) / Var(Rm)
SyllabusWeek Topic Applications Readings
1 Applications of reasoning econ, nlp, robotics, vision
2 Propositional reasoning verification, generation, plan GN87(2); Darwiche & Marquis
3 Bayesian Networks: semantics checking independence RN(13-14)
4Bayesian Networks: exact
inference vision; NLP QA RN(14); Amir'08; Pedro Feldenszwab
5 Bayesian Networks: learning NB generalizations Mitchell (stat-learning); RN(18?)
6 Sampling2-layer networks: diagnosis, state
est. Jordan ed.(MacKay ch)
7 Variational Approximations ? ?
8 DBNsLocalizing robots, Market
prediction Jordan book (KF ch), RN(15)
9 Logical Filtering Conveyor belt, Kriegspiel Amir'08 tech report on Logical Filtering
10 Logical Particle Filtering NLP, SLAM Hajishirzi-Amir'07,'08
11 FOL: semantics NLP, SitCalc, Golog GN87(3); Shoenfield (1)
12 FOL: reasoning QA from NLP GN87(4-5)
13 Relational Probabilistic Models social netsMLNs (Pedro Domingos), Braz-etal'07 (with me and Dan
Roth)
14 Description Logics / Semantic Web www services ? - semantics web conference?
15Cross-cutting: uncertain graphs,
beliefs logistics, networks Chang-Amir'07; Shirazi-Amir'08
16 Cross-cutting: convex optimization ? ?
High Dimensionality
• Many objects: books, people, computers
• Many properties: price, size, location
• Many relationships: sick(Person,Disease), FacebookFriend(P1,P2), before(W1,W2,D)
• Many questions
• Learning from examples, trials, observations
mini-Project Ideas
• Experimental comparison of – SAT solvers– FOL reasoners– Probabilistic inference approaches
• Applications of reasoning methods– Reasoning about STD spread
(highlyActive(P), std(D),sick(P,D), sex(P1,P2))– Stock-price prediction (small)
Project Ideas (1-3 people)
• Applications of reasoning methods:– Games: cards (Poker, Bridge, 21, Stratego)– Commonsense knowledge collection (2-people team)– Probabilistic commonsense (2-people team)– Question answering from text (CCG)– Stock-price prediction (interactions)
• Enhanced reasoning methods– Approximate theories by eliminating variables– Dynamic lifted probabilistic inference– Lifted probabilistic inference: (a) BNs, (b) Observsn.– Lifted probabilistic database queries– Learning probabilistic partially observed dynamic models– Parsing as probabilistic reasoning– Bayesian SVMs
Knowledge in Different Forms
• CYC, OpenMind, SUMO – Commonsense
• Ontologies – frame-based, semantic web
• Medical knowledge
• Diseases/symptoms networks
• Dynamic systems
• Specific applications: NLP, Databases
Reasoning Tasks
• A robot moving and manipulating the world– Track the environment and its body (actions)– Update its knowledge with new information
(sensors & communications)– Make timely decisions– Safe decisions– Take uncertainty into account– Learning and generalizing from knowledge
Example
• A robot moving and manipulating the world
Reasoner+
KnowledgeWorld
Sensoryinformation
Actions/Decisions
ReasoningAlgorithm
KB
Symbols toSensors
TasksMngr
Example Use of Reasoning 1
• Task: select an action to perform
• Logical KB: (a) Prove that KB entails move_fwd (e.g.,FOL)
(b) Find a model of KB that satisfies move_fwd (e.g., propositional logic)
• Probabilistic KB:– Find the probability of move_fwd (e.g., BNs)– Find an action that gives best utility (MDPs)
Example Use of Reasoning 2
• Task: find cause of error Err
• Logical KB: Abduction: Find an explanation Exp such that KB Exp logically entails Err
• Probabilistic KB:– Find the set of variable assignments that has
maximum posterior probability given Err
Knowledge Representation and Reasoning (KR&R)
• Two agents interacting– Sales and purchase agent– Collaboration to achieve a task– Information agent and user agent
Reasoning Agent 1+
Knowledge Base 1
Agent 2+
Knowledge Base 2Response
Request
Knowledge Representation and Reasoning (KR&R)
• Query answering:– Formal verification of digital circuits– Temporal verification of programs– Prediction and explanation
Human / SoftwareReasoning with
A Knowledge BaseAnswer
Query
Tractability of Reasoning
• More expressive languages require more time to reason with
Expressivity – Tractability tradeoff
• Compact representations not always more efficient for reasoning
• Reasoning with a complete model many times easier than reasoning with general knowledge in the same language
Summary: Why, When, How KR&R
• Reasoning with knowledge is good when we are not sure about knowledge or query.
• The language of KB is determined by the application:– Need for expressive language– Need for fast/accurate response
• Knowledge is entered by hand or learned
• Tasks for reasoning algorithms vary
Administrativia of this Class
• 4-5 homeworks• 3 units (undergrads / grads):
– Choice between small project and decreased homeworks
– Full homeworks require CS473/CS573 knowledge
• 4 units (grads)– Full homeworks + project
• A light final exam will be given
Project Choice
• 12th lec. (Sep 29): Project proposals due (~1-pages)
• 18th lec. (Oct 20): Progress Review due (~3 page)
• Final Project Submission (Dec 1): Projects due