Artificial Intelligence and Agents: A Definition and Exploration

1. CSCE 531Artificial IntelligenceCh.1 [P]: Artificial Intelligence and Agents Fall 2008 Marco Valtorta mgv@cse.sc.edu

2. Acknowledgment • The slides are based on the textbook [AIMA] and other sources, including other fine textbooks • The other textbooks I considered are: • David Poole, Alan Mackworth, and Randy Goebel. Computational Intelligence: A Logical Approach. Oxford, 1998 • A second edition (by Poole and Mackworth) is under development. Dr. Poole allowed us to use a draft of it in this course • Ivan Bratko. Prolog Programming for Artificial Intelligence, Third Edition. Addison-Wesley, 2001 • The fourth edition is under development • George F. Luger. Artificial Intelligence: Structures and Strategies for Complex Problem Solving, Sixth Edition. Addison-Welsey, 2009

3. Artificial Intelligence: a Definition • Artificial Intelligence is the synthesis and analysis of agents that act intelligently. • An agent is something that acts in an environment. • An agent that acts intelligently if: • its actions are appropriate for its goals and circumstances • it is flexible to changing environments and goals • it learns from experience • it makes appropriate choices given perceptual limitations and finite computation • Some agents are not computational • e.g., wind and rain eroding a landscape • Are all intelligent agents computational? • Open question!

4. Central Hypotheses of AI • Symbol-system hypothesis: • Reasoning is symbol manipulation • Attributed to Allan Newell (1927-1992) and Herbert Simon (1916-2001) • Church-Turing thesis: • Any symbol manipulation can be carried out on a Turing machine • Alonzo Church (1903-1995) • Alan Turing (1912-1954)

5. Agents and Environments

6. Sample Agent: Robot • actions: • movement, grippers, speech, facial expressions,. . . • observations: • vision, sonar, sound, speech recognition, gesture recognition,. . . • goals: • deliver food, rescue people, score goals, explore,. . . • past experiences: • effect of steering, slipperiness, how people move,. . . • prior knowledge: • what is important feature, categories of objects, what a sensor tell us,. . .

7. Example Agent: Teacher • actions: • present new concept, drill, give test, explain concept,. . . • observations: • test results, facial expressions, errors, focus,. . . • goals: • particular knowledge, skills, inquisitiveness, social skills,. . . • past experiences: • prior test results, effects of teaching strategies, . . . • prior knowledge: • subject material, teaching strategies,. . .

8. Example agent: Medical Doctor • actions: • operate, test, prescribe drugs, explain instructions,. . . • observations: • verbal symptoms, test results, visual appearance. . . • goals: • remove disease, relieve pain, increase life expectancy, reduce costs,. . . • past experiences: • treatment outcomes, effects of drugs, test results given symptoms. . . • prior knowledge: • possible diseases, symptoms, possible causal relationships. . .

9. Example Agent: User Interface • actions: • present information, ask user, find another information source, filter information, interrupt,. . . • observations: • users request, information retrieved, user feedback, facial expressions. . . • goals: • present information, maximize useful information, minimize irrelevant information, privacy,. . . • past experiences: • effect of presentation modes, reliability of information sources,. . . • prior knowledge: • information sources, presentation modalities. . .

10. The Role of Representation • Choosing a representation involves balancing conflicting objectives • Different tasks require different representations • Representations should be expressive (epistemologically adequate) and efficient (heuristically adequate)

11. Desiderata of Representations • We want a representation to be • rich enough to express the knowledge needed to solve the problem • Epistemologically adequate • as close to the problem as possible: compact, natural and maintainable • amenable to efficient computation: able to express features of the problem we can exploit for computational gain • Heuristically adequate • learnable from data and past experiences • able to trade off accuracy and computation time

12. Dimensions of Complexity • Modularity: • Flat, modular, or hierarchical • Representation: • Explicit states or features or objects and relations • Planning Horizon: • Static or finite stage or indefinite stage or infinite stage • Sensing Uncertainty: • Fully observable or partially observable • Process Uncertainty: • Deterministic or stochastic dynamics • Preference Dimension: • Goals or complex preferences • Number of agents: • Single-agent or multiple agents • Learning: • Knowledge is given or knowledge is learned from experience • Computational Limitations: • Perfect rationality or bounded rationality

13. Modularity • You can model the system at one level of abstraction: flat • Manuscript distinguish flat (no organizational structure) from modular (interacting modules that can be understood on their own; hierarchical seems to be a special case of modular) • You can model the system at multiple levels of abstraction: hierarchical • Example: Planning a trip from here to a resort in Cancun, Mexico • Flat representations are ok for simple systems, but complex biological systems, computer systems, organizations are all hierarchical • A flat description is either continuous or discrete. • Hierarchical reasoning is often a hybrid of continuous and discrete

14. Succinctness and Expressiveness • Much of modern AI is about finding compact representations and exploiting that compactness for computational gains. • An agent can reason in terms of: • explicit states • features or propositions. • It's often more natural to describe states in terms of features • 30 binary features can represent 230 = 1,073,741,824 states. • individuals and relations • There is a feature for each relationship on each tuple of individuals. • Often we can reason without knowing the individuals or when there are infinitely many individuals

15. Example: States Thermostat for a heater • 2 belief (i.e., internal) states: off, heating • 3 environment (i.e., external) states: cold, comfortable, hot • 6 total states corresponding to the different combinations of belief and environment states

16. Example: Features or Propositions Character recognition • Input is a binary image which is a 30x30 grid of pixels • Action is to determine which of the letters {a…z} is drawn in the image • There are 2900 different states of the image, and so 262900 different functions from the image state into the letters • We cannot even represent such functions in terms of the state space • Instead, we define features of the image, such as line segments, and define the function from images to characters in terms of these features

17. Example: Relational Descriptions University Registrar Agent • Propositional description: • “passed” feature for every student-course pair that depends on the grade feature for that pair • Relational description: • individual students and courses • relations grade and passed • Define how “passed” depends on grade once, and apply it for each student and course. Moreover this can be done before you know of any of the individuals, and so before you know the value of any of the features covers_core_courses(St, Dept) <- core_courses(dept, CC, MinPass) & passed_each(CC, St, MinPass). passed(St, C, MinPass) <- grade(St, C, Gr) & Gr >= MinPass.

18. Planning Horizon How far the agent looks into the future when deciding what to do • Static: world does not change • Finite stage: agent reasons about a fixed finite number of time steps • Indefinite stage: agent is reasoning about finite, but not predetermined, number of time steps • Infinite stage: the agent plans for going on forever (process oriented)

19. Uncertainty • There are two dimensions for uncertainty • Sensing uncertainty • Process uncertainty • In each dimension we can have • no uncertainty: the agent knows which world is true • disjunctive uncertainty: there is a set of worlds that are possible • probabilistic uncertainty: a probability distribution over the worlds

20. Uncertainty • Sensing uncertainty: Can the agent determine the state from the observations? • Fully-observable: the agent knows the state of the world from the observations. • Partially-observable: many states are possible given an observation. • Process uncertainty: If the agent knew the initial state and the action, could it predict the resulting state? • Deterministic dynamics: the state resulting from carrying out an action in state is determined from the action and the state • Stochastic dynamics: there is uncertainty over the states resulting from executing a given action in a given state.

21. Bounded Rationality Solution quality as a function of time for an anytime algorithm

22. Examples of Representational Frameworks • State-space search • Classical planning • Influence diagrams • Decision-theoretic planning • Reinforcement Learning

23. State-Space Search • flat or hierarchical • explicit states or features or objects and relations • static or finite stage or indefinite stage or infinite stage • fully observable or partially observable • deterministic or stochastic actions • goals or complex preferences • single agent or multiple agents • knowledge is given or learned • perfect rationality or bounded rationality

24. Classical Planning • flat or hierarchical • explicit states or features or objects and relations • static or finite stage or indefinite stage or infinite stage • fully observable or partially observable • deterministic or stochastic actions • goals or complex preferences • single agent or multiple agents • knowledge is given or learned • perfect rationality or bounded rationality

25. Influence Diagrams • flat or hierarchical • explicit states or features or objects and relations • static or finite stage or indefinite stage or infinite stage • fully observable or partially observable • deterministic or stochastic actions • goals or complex preferences • single agent or multiple agents • knowledge is given or learned • perfect rationality or bounded rationality

26. Decision-Theoretic Planning • flat or hierarchical • explicit states or features or objects and relations • static or finite stage or indefinite stage or infinite stage • fully observable or partially observable • deterministic or stochastic actions • goals or complex preferences • single agent or multiple agents • knowledge is given or learned • perfect rationality or bounded rationality

27. Reinforcement Learning • flat or hierarchical • explicit states or features or objects and relations • static or finite stage or indefinite stage or infinite stage • fully observable or partially observable • deterministic or stochastic actions • goals or complex preferences • single agent or multiple agents • knowledge is given or learned • perfect rationality or bounded rationality

28. Comparison of Some Representations

29. Four Application Domains • Autonomous delivery robot roams around an office environment and delivers coee, parcels, etc. • Diagnostic assistant helps a human troubleshootproblems and suggests repairs or treatments • E.g., electrical problems, medical diagnosis • Intelligent tutoring system teaches students in some subject area • Trading agent buys goods and services on your behalf

30. Environment for Delivery Robot

31. Example inputs: Prior knowledge: its capabilities, objects it may encounter, maps. Past experience: which actions are useful and when, what objects are there, how its actions aect its position Goals: what it needs to deliver and when, tradeoffs between acting quickly and acting safely Observations: about its environment from cameras, sonar, sound, laser range finders, or keyboards Sample activities: Determine where Craig's office is. Where coffee is, etc. Find a path between locations Plan how to carry out multiple tasks Make default assumptions about where Craig is Make tradeoffs under uncertainty: should it go near the stairs? Learn from experience. Sense the world, avoid obstacles, pickup and put down coffee Autonomous Delivery Robot

32. Environment for Diagnostic Assistant

33. Example inputs: Prior knowledge: how switches and lights work, how malfunctions manifest themselves, what information tests provide, the side effects of repairs Past experience: the effects of repairs or treatments, the prevalence of faults or diseases Goals: fixing the device and tradeoffs between fixing or replacing different components Observations: symptoms of a device or patient Sample activities: Derive the effects of faults and interventions Search through the space of possible fault complexes Explain its reasoning to the human who is using it Derive possible causes for symptoms; rule out other causes Plan courses of tests and treatments to address the problems Reason about the uncertainties/ambiguities given symptoms. Trade off alternate courses of action Learn what symptoms are associated with faults, the effects of treatments, and the accuracy of tests. Diagnostic Assistant

34. Example inputs: Prior knowledge: the ontology of what things are available, where to purchase items, how to decompose a complex item Past experience: how long special last, how long items take to sell out, who has good deals, what your competitors do Goals: what the person wants, their tradeoffs Observations: what items are available, prices, number in stock Sample activities: Trading agent interacts with an information environment to purchase goods and services. It acquires a users needs, desires and preferences. It finds what is available. It purchases goods and services that t together to fulfill user preferences. It is difficult because user preferences and what is available can change dynamically, and some items may be useless without other items. Trading Agent

35. Example inputs Prior knowledge: subject material, primitive strategies Past experience: common errors, effects of teaching strategies Goals: teach subject material, social skills, study skills, inquisitiveness, interest Observations: test results, facial expressions, questions, what the student is concentrating on Sample activities: To be filled Intelligent Tutoring Systems

36. Common tasks of the Domains • Modeling the environment: • Build models of the physical environment, patient, or information environment • Evidential reasoning or perception: • Given observations, determine what the world is like • Action: • Given a model of the world and a goal, determine what should be done • Learning from past experiences: • Learn about the specific case and the population of cases