1. Introduction to AI and Intelligent Agents Foundations of Artificial Intelligence

2. 2 Some Definitions of AI  Building systems that think like humans  “The exciting new effort to make computers think … machines with minds, in the full and literal sense” -- Haugeland, 1985  “The automation of activities that we associate with human thinking, … such as decision-making, problem solving, learning, …” -- Bellman, 1978  Building systems that act like humans  “The art of creating machines that perform functions that require intelligence when performed by people” -- Kurzweil, 1990  “The study of how to make computers do things at which, at the moment, people are better” -- Rich and Knight, 1991

3. Foundations of Artificial Intelligence 3 Some Definitions of AI  Building systems that think rationally  “The study of mental faculties through the use of computational models” -- Charniak and McDermott, 1985  “The study of the computations that make it possible to perceive, reason, and act” -- Winston, 1992  Building systems that act rationally  “A filed of study that seeks to explain and emulate intelligent behavior in terms of computational processes” -- Schalkoff, 1990  “The branch of computer science that is concerned with the automation of intelligent behavior” -- Luger and Stubblefield, 1993

4. Foundations of Artificial Intelligence 4 Thinking and Acting Humanly  Thinking humanly: cognitive modeling  Develop a precise theory of mind, through experimentation and introspection, then write a computer program that implements it  Example: GPS - General Problem Solver (Newell and Simon, 1961)  trying to model the human process of problem solving in general  Acting humanly  "If it looks, walks, and quacks like a duck, then it is a duck”  The Turing Test  interrogator communicates by typing at a terminal with TWO other agents. The human can say and ask whatever s/he likes, in natural language. If the human cannot decide which of the two agents is a human and which is a computer, then the computer has achieved AI  this is an OPERATIONAL definition of intelligence, i.e., one that gives an algorithm for testing objectively whether the definition is satisfied

5. Foundations of Artificial Intelligence 5 Thinking and Acting Rationally  Thinking Rationally  Capture ``correct'' reasoning processes”  A loose definition of rational thinking: Irrefutable reasoning process  How do we do this  Develop a formal model of reasoning (formal logic) that “always” leads to the “right” answer  Implement this model  How do we know when we've got it right?  when we can prove that the results of the programmed reasoning are correct  soundness and completeness of first-order logic  Acting Rationally  Act so that desired goals are achieved  The rational agent approach (this is what we’ll focus on in this course)  Figure out how to make correct decisions, which sometimes means thinking rationally and other times means having rational reflexes  correct inference versus rationality  reasoning versus acting; limited rationality

6. Turing’s Goal  Alan Turing, Computing Machinery and Intelligence, 1950:  Can machines think?  How could we tell? “I propose to consider the question, ‘Can machines think?’ This should begin with definitions of the meaning of the terms ‘machine’ and ‘think’. The definitions might be framed so as to reflect so far as possible the normal use of the words, but this attitude is dangerous. If the meaning of the words ‘machine’ and ‘think’ are to be found by examining how they are commonly used it is difficult to escape the conclusion that the meaning and the answer to the question, ‘Can machines think?’ is to be sought in a statistical survey such as a Gallup poll. But this is absurd. Instead of attempting such a definition I shall replace the question by another, which is closely related to it and is expressed in relatively unambiguous words.” — Alan Turing, Computing machinery and intelligence, 1950

7. Turing’s “Imitation Game” InterrogatorB (a person)A (a machine)

8. Necessary versus Sufficient Conditions  Is ability to pass a Turing Test a necessary condition of intelligence?  “May not machines carry out something which ought to be described as thinking but which is very different from what a man does? This objection is a very strong one, but at least we can say that if, nevertheless, a machine can be constructed to play the imitation game satisfactorily, we need not be troubled by this objection.” — Turing, 1950  Is ability to pass a Turing Test a sufficient condition of intelligence?

9. The Turing Syllogism  If an agent passes a Turing Test, then it produces a sensible sequence of verbal responses to a sequence of verbal stimuli.  If an agent produces a sensible sequence of verbal responses to a sequence of verbal stimuli, then it is intelligent.  Therefore, if an agent passes a Turing Test, then it is intelligent. The Capacity Conception: If an agent has the capacity to produce a sensible sequence of verbal responses to a sequence of verbal stimuli, whatever they may be, then it is intelligent.

10. Memorizing all possible answers? (Bertha’s Machine)

11. Foundations of Artificial Intelligence 11 Exponential Growth  Assume each time the judge asks a question, she picks between two questions based on what has happened so far Questions Asked Possible responses 12 24 38 416 532 664 n 2 n

12. Storage versus Length exponential

13. Foundations of Artificial Intelligence 13 n=10n=20n=30n=40n=50n=60 n.00001 second.00002 second.00003 second.00004 second.00005 second.00006 second 2n2n.001 second 1.0 second 17.9 minutes 12.7 days 35.7 years 336 centuries 3n3n.059 second 58 minutes 6.5 years 3855 centuries 2x10 8 centuries 1.3x10 13 centuries (one algorithm step = 1 microsecond) (Garvey & Johnson 1979) Polynomial vs. exponential time complexity

14. The Compact Conception  If an agent has the capacity to produce a sensible sequence of verbal responses to an arbitrary sequence of verbal stimuli without requiring exponential storage, then it is intelligent.

15. Size of the Universe Here, now Big bang 15*10 9 light-years Time

16. Storage Capacity of the Universe Volume: (15*10 9 light-years) 3 = (15*10 9 *10 16 meters) 3 Density: 1 bit per (10 -35 meters) 3 Total storage capacity: 10 184 bits < 10 200 bits < 2 670 bits Critical Turing Test length: 670 bits < 670 characters < 140 words < 1 minute The universe is not big enough to hold a bertha machine

17. Foundations of Artificial Intelligence 17 Some Sub-fields of AI  Problem solving  Lots of early success here  Solving puzzles  Playing chess  Mathematics (integration)  Uses techniques like search and problem reduction  Logical reasoning  Prove things by manipulating database of facts  Theorem proving  Automatic Programming  Writing computer programs given some sort of description  Some success with semi-automated methods  Some error detection systems  Automatic program verification

18. Foundations of Artificial Intelligence 18 Some Sub-fields of AI  Language understanding and semantic modeling  One of the earliest problems  Some success within limited domains  How can we “understand” written/spoken language?  Includes answering questions, translating between languages, learning from written text, and speech recognition  Some aspects of language understanding:  Associating spoken words with “actual” word  Understanding language forms, such as prefixes/suffixes/roots  Syntax; how to form grammatically correct sentences  Semantics; understanding meaning of words, phrases, sentences  Context  Conversation

19. Foundations of Artificial Intelligence 19 Some Sub-fields of AI  Pattern Recognition  Computer-aided identification of objects/shapes/sounds  Needed for speech and picture understanding  Requires signal acquisition, feature extraction,...  Data mining and Information Retrieval  Expert Systems and Knowledge-based Systems  Designers often called knowledge engineers  Translate things that an expert knows and rules that an expert uses to make decisions into a computer program  Problems include  Knowledge acquisition (or how do we get the information)  Explanation (of the answers)  Knowledge models (what do we do with info)  Handling uncertainty

20. Foundations of Artificial Intelligence 20 Some Sub-fields of AI  Planning, Robotics and Vision  Planning how to perform actions  Manipulating devices  Recognizing objects in pictures  Machine Learning and Neural Networks  Can we “remember” solutions, rather than recalculating them?  Can we learn additional facts from present data?  Can we model the physical aspects of the brain?  Classification and clustering  Non-monotonic Reasoning  Truth maintenance systems

21. Foundations of Artificial Intelligence 21 Fundamental Techniques of AI  Knowledge Representation  Intelligence/intelligent behavior requires knowledge, which is:  Voluminous  Hard to characterize  Constantly changing  How can one capture formally (i.e., computerize) everything needed for intelligent behavior? Some questions...  How do you store all of that data in a useful way?  Can you get rid of some?  How can you store decision making steps?  Characteristics of good data representation techniques:  Captures general situation rather than being overly specific  Understandable by the people who provide it  Easily modified to handle errors, changes in data, and changes in perception  Of general use

22. Foundations of Artificial Intelligence 22 Fundamental Techniques of AI  Search  How can we model the problem search space  How can we move between steps in a decision making process?  How can you find the info you need in a large data set?  Given a choice of possible decision sequences, how do you pick a good one?  Heuristic functions  Given a goal, how do you figure out what to do (planning)?  Base-level versus meta-level reasoning  How can we reason about what step to take next (in reaching the goal)?  How much do we reason before acting?

23. Foundations of Artificial Intelligence 23 AI in Everyday Life?  AI techniques are used in many common applications  Intelligent user interfaces  Search Engines  Spell/grammar checkers  Context sensitive help systems  Medical diagnosis systems  Regulating/Controlling hardware devices and processes (e.g, in automobiles)  Voice/image recognition (more generally, pattern recognition)  Scheduling systems (airlines, hotels, manufacturing)  Error detection/correction in electronic communication  Program verification / compiler and programming language design  Web search engines / Web spiders  Web personalization and Recommender systems (collaborative/content filtering)  Personal agents  Customer relationship management  Credit card verification in e-commerce / fraud detection  Data mining and knowledge discovery in databases  Computer games

24. Foundations of Artificial Intelligence 24 AI Spin-Offs  Many technologies widely used today were the direct or indirect results of research in AI:  The mouse  Time-sharing  Graphical user interfaces  Object-oriented programming  Computer games  Hypertext  Information Retrieval  The World Wide Web  Symbolic mathematical systems (e.g., Mathematica, Maple, etc.)  Very high-level programming languages  Web agents  Data Mining

25. Foundations of Artificial Intelligence 25 What is an Intelligent Agent  An agent is anything that can  perceive its environment through sensors, and  act upon that environment through actuators (or effectors)  Goal: Design rational agents that do a “good job” of acting in their environments  success determined based on some objective performance measure actuators

26. Foundations of Artificial Intelligence 26 Example: Vacuum Cleaner Agent  Percepts: location and contents, e.g., [A, Dirty]  Actions: Left, Right, Suck, NoOp

27. Foundations of Artificial Intelligence 27 What is an Intelligent Agent  Rational Agents  An agent should strive to "do the right thing", based on what it can perceive and the actions it can perform. The right action is the one that will cause the agent to be most successful.  Performance measure: An objective criterion for success of an agent's behavior.  E.g., performance measure of a vacuum-cleaner agent could be amount of dirt cleaned up, amount of time taken, amount of electricity consumed, amount of noise generated, etc.  Definition of Rational Agent:  For each possible percept sequence, a rational agent should select an action that is expected to maximize its performance measure, given the evidence provided by the percept sequence and whatever built-in knowledge the agent has.  Omniscience, learning, autonomy  Rationality is distinct from omniscience (all-knowing with infinite knowledge)  Choose action that maximizes expected value of perf. measure given percept to date  Agents can perform actions in order to modify future percepts so as to obtain useful information (information gathering, exploration)  An agent is autonomous if its behavior is determined by its own experience (with ability to learn and adapt)

28. Foundations of Artificial Intelligence 28 What is an Intelligent Agent  Rationality depends on  the performance measure that defines degree of success  the percept sequence - everything the agent has perceived so far  what the agent know about its environment  the actions that the agent can perform  Agent Function (percepts ==> actions)  Maps from percept histories to actions f: P *  A  The agent program runs on the physical architecture to produce the function f  agent = architecture + program Action := Function(Percept Sequence) If (Percept Sequence) then do Action  Example: A Simple Agent Function for Vacuum World If (current square is dirty) then suck Else move to adjacent square

29. Foundations of Artificial Intelligence 29 What is an Intelligent Agent  Limited Rationality  Optimal (i.e. best possible) rationality is NOT perfect success: limited sensors, actuators, and computing power may make this impossible  Theory of NP-completeness: some problems are likely impossible to solve quickly on ANY computer  Both natural and artificial intelligence are always limited  Degree of Rationality: the degree to which the agent’s internal "thinking" maximizes its performance measure, given  the available sensors  the available actuators  the available computing power  the available built-in knowledge

30. Foundations of Artificial Intelligence 30 PEAS Analysis  To design a rational agent, we must specify the task environment  PEAS Analysis:  Specify Performance Measure, Environment, Actuators, Sensors  Example: Consider the task of designing an automated taxi driver  Performance measure: Safe, fast, legal, comfortable trip, maximize profits  Environment: Roads, other traffic, pedestrians, customers  Actuators: Steering wheel, accelerator, brake, signal, horn  Sensors: Cameras, sonar, speedometer, GPS, odometer, engine sensors, keyboard

31. Foundations of Artificial Intelligence 31 PEAS Analysis – More Examples  Agent: Medical diagnosis system  Performance measure: Healthy patient, minimize costs, lawsuits  Environment: Patient, hospital, staff  Actuators: Screen display (questions, tests, diagnoses, treatments, referrals)  Sensors: Keyboard (entry of symptoms, findings, patient's answers)  Agent: Part-picking robot  Performance measure: Percentage of parts in correct bins  Environment: Conveyor belt with parts, bins  Actuators: Jointed arm and hand  Sensors: Camera, joint angle sensors

32. Foundations of Artificial Intelligence 32 PEAS Analysis – More Examples  Agent: Internet Shopping Agent  Performance measure??  Environment??  Actuators??  Sensors??

33. Foundations of Artificial Intelligence 33 Environment Types  Fully observable (vs. partially observable):  An agent's sensors give it access to the complete state of the environment at each point in time.  Deterministic (vs. stochastic):  The next state of the environment is completely determined by the current state and the action executed by the agent. (If the environment is deterministic except for the actions of other agents, then the environment is strategic).  Episodic (vs. sequential):  The agent's experience is divided into atomic "episodes" (each episode consists of the agent perceiving and then performing a single action), and the choice of action in each episode depends only on the episode itself.

34. Foundations of Artificial Intelligence 34 Environment Types (cont.)  Static (vs. dynamic):  The environment is unchanged while an agent is deliberating (the environment is semi-dynamic if the environment itself does not change with the passage of time but the agent's performance score does).  Discrete (vs. continuous):  A limited number of distinct, clearly defined percepts and actions.  Single agent (vs. multi-agent):  An agent operating by itself in an environment.

35. Foundations of Artificial Intelligence 35 Environment Types (cont.) The environment type largely determines the agent design. The real world is (of course) partially observable, stochastic, sequential, dynamic, continuous, multi-agent

36. Foundations of Artificial Intelligence 36 Structure of an Intelligent Agent  All agents have the same basic structure:  accept percepts from environment  generate actions  A Skeleton Agent:  Observations:  agent may or may not build percept sequence in memory (depends on domain)  performance measure is not part of the agent; it is applied externally to judge the success of the agent function Skeleton-Agent(percept) returns action static: memory, the agent's memory of the world memory  Update-Memory(memory, percept) action  Choose-Best-Action(memory) memory  Update-Memory(memory, action) return action function Skeleton-Agent(percept) returns action static: memory, the agent's memory of the world memory  Update-Memory(memory, percept) action  Choose-Best-Action(memory) memory  Update-Memory(memory, action) return action

37. Foundations of Artificial Intelligence 37 Looking Up the Answer?  A Template for a Table-Driven Agent:  Why can't we just look up the answers?  The disadvantages of this architecture  infeasibility (excessive size)  lack of adaptiveness  How big would the table have to be?  Could the agent ever learn from its mistakes?  Where should the table come from in the first place? function Table-Driven-Agent(percept) returns action static: percepts, a sequence, initially empty table, a table indexed by percept sequences, initially fully specified append percept to the end of percepts action  LookUp(percepts, table) return action function Table-Driven-Agent(percept) returns action static: percepts, a sequence, initially empty table, a table indexed by percept sequences, initially fully specified append percept to the end of percepts action  LookUp(percepts, table) return action

38. Foundations of Artificial Intelligence 38 Agent Types  Simple reflex agents  are based on condition-action rules and implemented with an appropriate production system. They are stateless devices which do not have memory of past world states.  Reflex Agents with memory (Model-Based)  have internal state which is used to keep track of past states of the world.  Agents with goals  are agents which in addition to state information have a kind of goal information which describes desirable situations. Agents of this kind take future events into consideration.  Utility-based agents  base their decision on classic axiomatic utility-theory in order to act rationally. Note: All of these can be turned into “learning” agents

39. Foundations of Artificial Intelligence 39 A Simple Reflex Agent function Simple-Reflex-Agent(percept) returns action static: rules, a set of condition-action rules state  Interpret-Input(percept) rule  Rule-Match(state, rules) action  Rule-Action[rule] return action function Simple-Reflex-Agent(percept) returns action static: rules, a set of condition-action rules state  Interpret-Input(percept) rule  Rule-Match(state, rules) action  Rule-Action[rule] return action  We can summarize part of the table by formulating commonly occurring patterns as condition-action rules:  Example: if car-in-front-brakes then initiate braking  Agent works by finding a rule whose condition matches the current situation  rule-based systems  But, this only works if the current percept is sufficient for making the correct decision

40. Foundations of Artificial Intelligence 40 Example: Simple Reflex Vacuum Agent

41. Foundations of Artificial Intelligence 41 Agents that Keep Track of the World function Reflex-Agent-With-State(percept) returns action static: rules, a set of condition-action rules state, a description of the current world state  Update-State(state, percept) rule  Rule-Match(state, rules) action  Rule-Action[rule] state  Update-State(state, action) return action function Reflex-Agent-With-State(percept) returns action static: rules, a set of condition-action rules state, a description of the current world state  Update-State(state, percept) rule  Rule-Match(state, rules) action  Rule-Action[rule] state  Update-State(state, action) return action  Updating internal state requires two kinds of encoded knowledge  knowledge about how the world changes (independent of the agents’ actions)  knowledge about how the agents’ actions affect the world  But, knowledge of the internal state is not always enough  how to choose among alternative decision paths (e.g., where should the car go at an intersection)?  Requires knowledge of the goal to be achieved

42. Foundations of Artificial Intelligence 42 Agents with Explicit Goals  Reasoning about actions  reflex agents only act based on pre-computed knowledge (rules)  goal-based (planning) act by reasoning about which actions achieve the goal  less efficient, but more adaptive and flexible

43. Foundations of Artificial Intelligence 43 Agents with Explicit Goals  Knowing current state is not always enough.  State allows an agent to keep track of unseen parts of the world, but the agent must update state based on knowledge of changes in the world and of effects of own actions.  Goal = description of desired situation  Examples:  Decision to change lanes depends on a goal to go somewhere (and other factors);  Decision to put an item in shopping basket depends on a shopping list, map of store, knowledge of menu  Notes:  Search (Russell Chapters 3-5) and Planning (Chapters 11-13) are concerned with finding sequences of actions to satisfy a goal.  Reflexive agent concerned with one action at a time.  Classical Planning: finding a sequence of actions that achieves a goal.  Contrast with condition-action rules: involves consideration of future "what will happen if I do..." (fundamental difference).

44. Foundations of Artificial Intelligence 44 A Complete Utility-Based Agent  Utility Function  a mapping of states onto real numbers  allows rational decisions in two kinds of situations  evaluation of the tradeoffs among conflicting goals  evaluation of competing goals

45. Foundations of Artificial Intelligence 45 Utility-Based Agents (Cont.)  Preferred world state has higher utility for agent = quality of being useful  Examples  quicker, safer, more reliable ways to get where going;  price comparison shopping  bidding on items in an auction  evaluating bids in an auction  Utility function: state ==> U(state) = measure of happiness  Search (goal-based) vs. games (utilities).

46. Foundations of Artificial Intelligence 46 Shopping Agent Example  Navigating: Move around store; avoid obstacles  Reflex agent: store map precompiled.  Goal-based agent: create an internal map, reason explicitly about it, use signs and adapt to changes (e.g., specials at the ends of aisles).  Gathering: Find and put into cart groceries it wants, need to induce objects from percepts.  Reflex agent: wander and grab items that look good.  Goal-based agent: shopping list.  Menu-planning: Generate shopping list, modify list if store is out of some item.  Goal-based agent: required; what happens when a needed item is not there? Achieve the goal some other way. e.g., no milk cartons: get canned milk or powdered milk.  Choosing among alternative brands  utility-based agent: trade off quality for price.

47. Foundations of Artificial Intelligence 47 General Architecture for Goal-Based Agents  Simple agents do not have access to their own performance measure  In this case the designer will "hard wire" a goal for the agent, i.e. the designer will choose the goal and build it into the agent  Similarly, unintelligent agents cannot formulate their own problem  this formulation must be built-in also  The while loop above is the "execution phase" of this agent's behavior  Note that this architecture assumes that the execution phase does not require monitoring of the environment Input percept state  Update-State(state, percept) goal  Formulate-Goal(state, perf-measure) search-space  Formulate-Problem (state, goal) plan  Search(search-space, goal) while (plan not empty) do action  Recommendation(plan, state) plan  Remainder(plan, state) output action end Input percept state  Update-State(state, percept) goal  Formulate-Goal(state, perf-measure) search-space  Formulate-Problem (state, goal) plan  Search(search-space, goal) while (plan not empty) do action  Recommendation(plan, state) plan  Remainder(plan, state) output action end

48. Foundations of Artificial Intelligence 48 Learning Agents  Four main components:  Performance element: the agent function  Learning element: responsible for making improvements by observing performance  Critic: gives feedback to learning element by measuring agent’s performance  Problem generator: suggest other possible courses of actions (exploration)

49. Foundations of Artificial Intelligence 49 Search and Knowledge Representation  Goal-based and utility-based agents require representation of:  states within the environment  actions and effects (effect of an action is transition from the current state to another state)  goals  utilities  Problems can often be formulated as a search problem  to satisfy a goal, agent must find a sequence of actions (a path in the state-space graph) from the starting state to a goal state.  To do this efficiently, agents must have the ability to reason with their knowledge about the world and the problem domain  which path to follow (which action to choose from) next  how to determine if a goal state is reached OR how decide if a satisfactory state has been reached.

50. Foundations of Artificial Intelligence 50 Intelligent Agent Summary  An agent perceives and acts in an environment. It has an architecture and is implemented by a program.  An ideal agent always chooses the action which maximizes its expected performance, given the percept sequence received so far.  An autonomous agent uses its own experience rather than built-in knowledge of the environment by the designer.  An agent program maps from a percept to an action and updates its internal state.  Reflex agents respond immediately to percepts.  Goal-based agents act in order to achieve their goal(s).  Utility-based agents maximize their own utility function.

51. Foundations of Artificial Intelligence 51 Exercise  Do Exercise 1.3, on Page 30  You can find out about the Loebner Prize at: http://www.loebner.net/Prizef/loebner-prize.html  Also (for discussion) look at exercise 1.2 and read the material on the Turing Test at: http://plato.stanford.edu/entries/turing-test/  Read the article by Jennings and Wooldridge (“Applications of Intelligent Agents”). Compare and contrast the definitions of agents and intelligent agents as given by Russell and Norvig (in the text book) and and in the article.

52. Foundations of Artificial Intelligence 52 Exercise  News Filtering Internet Agent  uses a static user profile (e.g., a set of keywords specified by the user)  on a regular basis, searches a specified news site (e.g., Reuters or AP) for news stories that match the user profile  can search through the site by following links from page to page  presents a set of links to the matching stories that have not been read before (matching based on the number of words from the profile occurring in the news story)  (1) Give a detailed PEAS description for the news filtering agent  (2) Characterize the environment type (as being observable, deterministic, episodic, static, etc).