Intelligent AgentsAIMA Chapter 2, 2nd Ed. (after Russell and Norvig)AIMA Chapter 2, 2nd Ed. (after Russell and Norvig) 1Outline♦ Agents and Environments♦ Rationality♦ Environment Specification and Types♦ Agent Functions, Programs, and TypesAIMA Chapter 2, 2nd Ed. (after Russell and Norvig) 2Agents Interact with EnvironmentsMust first specify the setting for intelligent agent designAn agent perceives its environment through sensors and acts upon itthrough actuatorsAIMA Chapter 2, 2nd Ed. (after Russell and Norvig) 3Example Sensors and ActuatorsHumans??Robots??Softbots??AIMA Chapter 2, 2nd Ed. (after Russell and Norvig) 4Example Sensors and ActuatorsHumans?? eyes and ears / hands and legsRobot?? cameras / motorsSoftbot?? keystrokes /displaysAIMA Chapter 2, 2nd Ed. (after Russell and Norvig) 5Agents and Environments (cont.)Mathematically, an agent function maps a ny percept sequence to an action(and thus describes beha vior)• percepts: agent’s perceptual inputs at any instance• percept sequence: complete history• action: an agent’s action choice at any instant can depend on the entirepercept sequenceProblematic from an implementation perspective (why?), so need agent pro-gramsAIMA Chapter 2, 2nd Ed. (after Russell and Norvig) 6Examples (cont.)Consider the task of designing an au tomated taxi:Percepts??Actions??Environment??AIMA Chapter 2, 2nd Ed. (after Russell and Norvig) 7Examples (cont.)Consider the task of designing an au tomated taxi:Percepts?? video, accelerometers, gauges, engine sensors, keyb oard, GPS,. . .Actions?? steer, accelerate, brake, horn, speak/display, . . .Environment?? US urban streets, freeways, traffic, ped estrians, weather,customers, . . .AIMA Chapter 2, 2nd Ed. (after Russell and Norvig) 8Another Example: Vacuum WorldPercepts??Actions??Environment??AIMA Chapter 2, 2nd Ed. (after Russell and Norvig) 9Another Example: Vacuum WorldPercepts?? location, dirtinessActions?? suck, left, right, no-opEnvironment?? grid, walls/obstacles, dirt distribu tion and creation, agentbody (movement actions work unless b ump into wall, suck actions pu t dirtinto agent body (or not))AIMA Chapter 2, 2nd Ed. (after Russell and Norvig) 10Simple Agent Function for Vacuum WorldPartial tabulation of this simple agent functionPercept sequence Action(A, Clean) Right(A, Dirty) Suck(B, Clean) Left(B, Dirty) Suck(A, Clean) (A, Clean) Right. . . . . .How can we define different vacuum world agents?What is the obvious que stion for AI?AIMA Chapter 2, 2nd Ed. (after Russell and Norvig) 11Agent ProgramAgent function: If the current square is dirty, then suck dirt; otherwise, moveto the other square.function Reflex-Vacuum-Agent( [location,status]) returns an actionif status = Dirty then return Suckelse if location = A then return Rightelse if location = B then return LeftAIMA Chapter 2, 2nd Ed. (after Russell and Norvig) 12Good Behavior: RationalityA rational agent is one that does “the right thing”, e.g., every entry in theaction function table is filled out correctly• the right action is the one that will cause the agent to be most successful• therefore, we n e ed to be able to measure success• a performance measure embodies th e criterion for success of an agent’sbehaviorAIMA Chapter 2, 2nd Ed. (after Russell and Norvig) 13Performance MeasuresPerformance measure: an objective, numerical value for any environmenthistoryWhat are reasonable performance measures for the vacuum world?AIMA Chapter 2, 2nd Ed. (after Russell and Norvig) 14Performance MeasuresPerformance measure: an objective, numerical value for any environmenthistoryWhat are reasonable performance measures for the vacuum world?• the amount of dirt cleaned up in an hour– one point per square clea ned up in time T ?– one point per clean square per time step, minus one per move?– penalize for > k dirty squares?• having a clean floor• generally better to measure what you want in the environment, ratherthan how you think the agent should behave• difficult to come up with measures (sustained mediocrity vs. high s andlows)AIMA Chapter 2, 2nd Ed. (after Russell and Norvig) 15RationalityRationality depend s on• the performance measure defining the success criterion• the agent’s prior knowledge of the environment• the actions that the agent can perform• the agent’s perce pt sequence to dateRational action: whichever action maximizes the expected value of the per-formance measure given the percept sequence to date and built-in knowledgeRational agent: for each possible percept sequence, selects an action tha t isexpected to maximize its performance measureRational 6= omniscientRational 6= cla irvoyantRational 6= successfulAIMA Chapter 2, 2nd Ed. (after Russell and Norvig) 16Omniscience, Learning, and AutonomyRational 6= omniscient• airplane flattens p e rson crossing street example• rationality maximizes expected performance, depending on knowledge todate; perfection maximizes actual performance• crossing without looking is n ot rational because lacks information gath-ering (doing actions to modify future percepts, explor ation)Rational agents should also• learn from percepts (to augment or modify prior knowledge)• learn to be autonomous (rely on percepts rather than prior – often par tialand/or incorrect – knowledge)Rational ⇒ exploration, learning, autonomyAIMA Chapter 2, 2nd Ed. (after Russell and Norvig) 17Specifying the Task Environment: PEASTask Environments: “problems” to which “agents” are solutionsWe thus need to specify the problem before we develop the solutionExample: PEAS Specification for an Automated Taxi Driver Agent• Performance Measures: corr ec t destination, safe, fast, legal, comfortable,profitable, . . .• Environment: roads, tr affic, pedestrians, customers, . . .• Actuators: steering, accelerator, brake, horn, . . .• Sensors: camera, sonar, speedometer, GPS, . . .AIMA Chapter 2, 2nd Ed. (after Russell and Norvig) 18More PEAS ExamplesText-based C onversational Tutor• performance: maximize test score• environment: students, testing agency• actuators: display exercise, suggestions, corrections• sensors: keyboard entryWhat about a Speech-based Conversational Tutor?See Figure 2.5 for more examplesNOTE: toy 6= artificial environmentAIMA Chapter 2, 2nd Ed. (after Russell and Norvig) 19Environment DimensionsFu lly ver sus Partially Observable• fully is with respect to observation relevance for action choice (thus de-pends on