Plans for Today

1. Plans for Today • Chapter 2: Intelligent Agents (until break) • Lisp: Some questions that came up in lab • Resume intelligent agents after Lisp issues

2. Current Class Numbers • By my roster: • 3a: 30 students • 4a: 15 students

3. Intelligent Agents • Agent: anything that can be viewed as… • perceiving its environment through sensors • acting upon its environment through effectors • Diagram • Examples: • Human • Web search agent • Chess player • What are sensors and effectors for each of these?

4. Rational Agents • Rational Agent: one that does the right thing • Criteria: Performance measure • Performance measures for • Web search engine? • Tic-tac-toe player? Chess player? • How does when performance is measured play a role? • short vs. long term

5. Rational Agents • Omniscient agent • Knows the actual outcome of its actions • What information would a chess player need to have to be omniscient? • Omniscience is (generally) impossible • A rational agent should do the right thing based on the knowledge it has

6. Rational Agents • What is rational depends on four things: • Performance measure • Percept sequence: everything agent has seen so far • Knowledge agent has about environment • Actions agent is capable of performing • Ideal Rational Agent • Does whatever action is expected to maximize its performance measure, based on percept sequence and built-in knowledge

7. Ideal Mapping • Percept sequence is only varying element of rationality • Percept Sequence  Action • In theory, can build a table mapping percept sequence to action • Sample table for chess board • Ideal mapping describes behavior of ideal agents

8. The Mapping Table • In most cases, mapping is explosively too large to write down explicitly • In some cases, mapping can be defined via a specification • Example: agent to sort a list of numbers • Sample table for such an agent • Lisp code

9. Autonomy • “Independence” • A system is autonomous if its behavior is determined by its own experience • An alarm that goes off at a prespecified time is not autonomous • An alarm that goes off when smoke is sensed is autonomous • A system without autonomy lacks flexibility

10. Structure of Intelligent Agents • AI designes the agent program • The program runs on some kind of architecture • To design an agent program, need to understand • Percepts • Actions • Goals • Environment

11. Some Sample Agents • What are percepts, actions, goals, and environment for: • Chess Player? • Web Search Tool? • Matchmaker? • Musical performer? • Environments can be real (web search tool) or artificial (Turing test) • distinction is about whether it is a simulation for agent or the “real thing”

12. Agent Programs • Some extra Lisp: Persistence of state (static variables) • Allows a function to keep track of a variable over repeated calls. • Put functions inside a let block • (let ((sum 0)) (defun myfun (x) (setf sum (+ sum x))) (defun report () sum))

13. Skeleton Agent • Inputs: percept • Outputs: action • Static Variable: memory • Pseudocode: • memory  Update-Memory(memory, percept) • action  Choose-Best-Action(memory) • memory  Update-Memory(memory, action) • return action

14. Where we’re going today • Frameworks of Agent Programs • Types of Agent Programs • Types of Environments • Assignment 2 and associated code

15. Skeleton Agent • General purpose agent code • (let ((memory nil)) (defun skeleton-agent (percept) (setf memory (update-memory memory percept)) (setf action (choose-best-action memory)) (setf memory (update-memory memory action)) action ; return action ))

16. Table Lookup Agent • Inputs: percept • Outputs: action • Static Variable: percepts, table • Pseudocode: • Append percept to the end of percepts • Action  Lookup(percepts, table) • return action

17. Lookup Table Agent • General purpose agent code • (let ((percepts nil) (table ????) (defun table-lookup-agent (percept) (setf percepts (append (list percept) percepts)) (lookup percepts table)) ))

18. Specific Agent Example:Pathfinder (Mars Explorer) • Percepts: • Actions: • Goals: • Environment: • Would table-driven work? • Table would take massive memory to store • Long time for programmer to build table • No autonomy

19. Four kinds of better agent programs • Simple reflex agents • Agents that keep track of the world • Goal-based agents • Utility-based agents

20. Simple reflex agents • Specific response to percepts, i.e. condition-action rule • if new-boulder-in-sight thenmove-towards-new-boulder • Pretty easy to implement • Requires no knowledge of past • Limited what you can do with it

22. Agents that keep track of the world • Maintain an internal state which is adjusted by each percept • Example: • Internal state: looking for a new boulder, or rolling towards one • Affects how Pathfinder will react when seeing a new boulder • Rule for action depends on both state and percept • Different from reflex, which only depends on percept

24. Goal-Based Agents • Agent continues to receive percepts and maintain state • Agent also has a goal • Makes decisions based on achieving goal • Example • Pathfinder goal: reach a boulder • If pathfinder trips or gets stuck, can make decisions to reach goal

26. Utility-Based Agents • Goals are not enough – need to know value of goal • Is this a minor accomplishment, or a major one? • Affects decision making – will take greater risks for more major goals • Utility: numerical measurement of importance of a goal • A utility-based agent will attempt to make the appropriate tradeoff

28. Environments: Accessible vs. Inaccessible • Accessible: agents sensors detect all aspects of environment relevant to deciding action • Inaccessible: some relevant information is not available • Examples? • Which is more desirable?

29. Environments: Determinstic vs. Nondeterminstic • Deterministic: next state of environment is completely determined by current state and agent actions • Nondeterminstic: uncertainty as to next state • If environment is inaccessible but deterministic, may appear nondeterministic • Agent’s point of view is the important one • Examples? • Which is more desirable?

30. Environments: Episodic vs. Nonepisodic • Episodic: Experience is divided into “episodes” of agent perceiving than acting • Quality of action depends just on the episode itself, not on previous • Examples? • Which is more desirable?

31. Environments: Static vs. Dynamic • Dynamic: Environment can change while agent is thinking • Static: Environment does not change while agent thinks • Semidynamic: Environment does not change with time, but performance score does • Examples? • Which is more desirable?

32. Environments: Discrete vs. Continuous • Discrete: Percepts and actions are distinct, clearly defined, and often limited in number • Examples? • Which is more desirable?

33. Lisp Questions

34. Why the dot in cons?Two Explanations: • High level: cons expects a list in the second position • Lower level: • Cons takes a cons cell from the free storage list • Puts first argument in “first” position • Puts second argument in “rest” position • Separates by a dot, unless “rest” position is a pointer (indicates continuing list)

35. How does append work? • Makes copy of first list • Takes last pointer and points to second list • Picture

36. How to debug? • Can trace function calls with (trace function) and (untrace function) • Demonstration with mystery function from lab • At a Break> prompt, can see call stack with • backtrace • Can go through code step by step • (step (mystery 2 3)) • Use step and next to go through each function as you go along • Use (print var)

37. Random bits • Are Lisp functions pass by value or pass by reference? • Lisp does pass by value, but all lists are referenced by a pointer. So like C++, can affect original list this way. • Why the p in (zerop x)? • p = predicate • NOT true that p = positive

38. Scoping and binding • let declares a scope where variable bindings are insulated from outside • usual notions of local and global variables apply • if you want to change a global variable from within a function

39. Assignment 2: Vacuum-World • Demonstration • Percepts: Vacuum cleaner can only detect: • Bumped into something? • Dirt underneath me? • At home location? • Actions: forward, turn right, turn left, suck up dirt, turn off

40. Vacuum-World cont. • Goals: Clean up as much dirt as possible with as few moves as possible. • Specifically: • 100 points for each piece of dirt vacuumed • -1 point for each action • -1000 if it turns off when not at home • You will modify choose-action function • possibly others? • New Lisp: try out each function, explain as you go