Problem-Solving: A* and Beyond. CSCI 3202, Fall 2010

Transcription

1 Problem-Solving: A* and Beyond CSCI 3202, Fall 2010

2 Administrivia Reading for this week: (up to p. 126) in R+N; (pp ) Problem Set 1 will be posted by tomorrow morning; due September 23 in class. HARD COPY ONLY, please.

3 The Basic Idea Behind A* Search 1) Devise an "estimation" function, h, that gives you a plausible underestimate of the distance from a given state to the goal. 2) Call the cost of getting from the initial state to the current state g. 3) To measure the quality of a given state s, sum the cost (g) of arriving at this state and the estimate (h) of the cost required to get from s to the goal. 4) Use best-first search, where g+h is the overall quality function.

4 How do you estimate distance to the goal? (Or: What s that h function?) Try to find a metric that will be a close (informative) one,but that won t overestimate (it should be optimistic). Example: 8-9 puzzle Number of misplaced tiles Sum of Manhattan distance of misplaced tiles (better) Best solution length for (say) tiles 1-4 Example: Choosing a path on a map from city A to B Euclidean distance to B.

5 AStar-Best-First-Search [Nodes] IF there are no more nodes THEN FAIL ThisNode <-- First(Nodes) IF ThisNode is the goal THEN return ThisNode ELSE ChildNodes <-- Children(ThisNode) Astar-Best-First-Search (Sort-by-Lowest-Cost-plus-Estimate ChildNodes Remove-Node(ThisNode Nodes))

6 A few handy heuristics Redefine the edges (moves) of the problem graph, as by chunking (e.g., Rubik s cube) Eliminate chunks of the problem graph, as by constraints (e.g., SEND+MORE = MONEY problem) Decompose the problem into subproblems (Rubik s cube again; or Rush Hour)

7 Means-Ends Analysis A technique for finding the next step to take in a search process. Here, we find the most important difference between our current state and the goal state and use that difference (recursively) to dictate our next move.

8 To SOLVE a problem: If the current state is the goal state, then we re done. If not, first REDUCE the difference between the current state and goal state to produce a new current state. Then SOLVE the problem from the (presumably better) current state and the original goal. To REDUCE the difference between the current state and the goal state, find the most important difference between the current state and the goal state. Find an operator that tends to reduce that difference, and APPLY that operator to the current state. To APPLY an operator to the current state, see if there are any differences between the current state and any preconditions for the operator. If there are, REDUCE the difference between the current state and the preconditions for the operator. If not return the result of APPLYing the operator directly to the current state.

9 Local search Up until now, we ve imagined a two-phase method: find a solution path (search the problem space graph), and then return that path so that it can be executed. In local search, we simply search the graph directly, hoping to land on a goal state. Two simple methods: hill-climbing and beam search.

10 Hill-Climb [Current-Node] IF Current-Node is the goal THEN return Current-Node ELSE ChildNodes Children (Current-Node) BestChild LooksBest (ChildNodes) Hill-Climb [BestChild] Beam-Search [Set-of-Nodes] If Set-of-Nodes contains a goal state THEN return that state ELSE AllChildNodes Find-All-Children (Set-of-Nodes) BestNChildren N-Best-Looking (AllChildNodes) Beam-Search (BestNChildren)

11 An Example of Hill-Climbing: the Word Arithmetic Problem Let each node of the problem space graph be a candidate solution: SD*ENM*OYR Each edge from this node leads to a node in which two assignments have been switched: SY*ENM*ODR

12 Hill-Climbing Example, Continued Now we hill-climb by starting at some candidate solution and moving to the adjacent solution that causes the fewest arithmetic errors.

13 Problem-Solving: The Cognitive Side Do we actually use (say) depth-first search? Or breadth-first search? Or A*? Is the problem-space formalism useful,or natural? How do we solve problems, anyhow?

14 "A man gets an unsigned letter telling him to go to the local graveyard at midnight. He does not generally pay attention to such things, but complies out of curiosity. It is a deathly still night, lighted by a thin crescent moon. The man stations himself in front of his family's ancestral crypt. The man is about to leave when he hears scraping footsteps. He yells out, but no one answers. The next morning, the caretaker finds the man dead in front of the crypt, a hideous grin on his face. Did the man vote for Teddy Roosevelt in the 1904 U.S. presidential election?" -- Poundstone

15 Hanging from the ceiling are two lengths of string, each six feet long. The strings are ten feet apart. As the room is ten feet high, the lower ends of the strings hang four feet above the floor. On the floor is a Swiss Army knife, a firecracker, a small glass of water, a twenty-five pound block of ice in an aluminum pan, and a friendly Labrador retriever. Your job is to tie the two ends of the two strings together. -- based on Poundstone, "Labyrinths of Reason"

16 All astronomers are stamp collectors. No baseball players are astronomers.???

17 Other ways of looking at the "search" issue; "insight" problems Retrieving schemas/searching for patterns Restructuring the problem space itself Functional fixedness and the "Einstellung" effect

18 Two Physics Problems Rolling a quarter around another quarter... Elastic collision (ping-pong vs. bowling ball)

19 Mr. Smith and his wife invited four other couples for a party. When everyone arrived, some of the people in the room shook hands with some of the others. Of course, nobody shook hands with their spouse and nobody shook hands with the same person twice. After that, Mr. Smith asked everyone how many times they shook someone s hand. He received different answers from everybody. How many times did Mrs. Smith shake someone s hand? From Michalewicz and Fogel, How to Solve It: Modern Heuristics

20 Game-Playing: the Essential Arsenal The idea of a minimax search Avoiding useless search: the alpha-beta algorithm Other ways of pruning the search tree Other central issues in game-playing

21 The Game of Sim

22 In Scheme: (define (minimax gamestate myfunction otherfunction level) (cond ((= level 0) (getvalue gamestate)) (else (let* ((successors (getnextmoves gamestate)) (othervals (map (lambda (successor) (minimax successor otherfunction myfunction (- level 1))) successors))) (apply myfunction othervals)))))

23 (define (alpha-beta alpha beta level node max?) (cond ((= level 0) (static-value node)) (max? (let ((children (find-next-moves node #t))) (max-look-at-each-child-node children alpha beta level))) (else (let ((children (find-next-moves node #f))) (min-look-at-each-child-node children alpha beta level))))) (define (max-look-at-each-child-node children alpha beta level) (cond ((null? children) alpha) ((>= alpha beta) alpha) (else (let ((next-value (alpha-beta alpha beta (- level 1) (first children) #f))) (max-look-at-each-child-node (rest children) (max next-value alpha) beta level))))) (define (min-look-at-each-child-node children alpha beta level) (cond ((null? children) beta) ((>= alpha beta) beta) (else (let ((next-value (alpha-beta alpha beta (- level 1) (first children) #t))) (min-look-at-each-child-node (rest children) alpha (min next-value beta) level)))))

24 Calling alpha-beta with starting values of a (very low) alpha value, a (very high) beta value, a depth of 5, the starting game state, and from the point of view of the maximizer. (alpha-beta initial-game-state #t)