DOC PREVIEW
CORNELL CS 4700 - Study Notes

This preview shows page 1-2-17-18-19-36-37 out of 37 pages.

Save
View full document
View full document
Premium Document
Do you want full access? Go Premium and unlock all 37 pages.
Access to all documents
Download any document
Ad free experience
View full document
Premium Document
Do you want full access? Go Premium and unlock all 37 pages.
Access to all documents
Download any document
Ad free experience
View full document
Premium Document
Do you want full access? Go Premium and unlock all 37 pages.
Access to all documents
Download any document
Ad free experience
View full document
Premium Document
Do you want full access? Go Premium and unlock all 37 pages.
Access to all documents
Download any document
Ad free experience
View full document
Premium Document
Do you want full access? Go Premium and unlock all 37 pages.
Access to all documents
Download any document
Ad free experience
View full document
Premium Document
Do you want full access? Go Premium and unlock all 37 pages.
Access to all documents
Download any document
Ad free experience
View full document
Premium Document
Do you want full access? Go Premium and unlock all 37 pages.
Access to all documents
Download any document
Ad free experience
Premium Document
Do you want full access? Go Premium and unlock all 37 pages.
Access to all documents
Download any document
Ad free experience

Unformatted text preview:

Informed SearchCombinatorial Explosion Depth Nodes Time Memory 2 1100 .11 sec 1 meg 4 111,100 11 sec 106 meg 6 107 19 min 10 gig 8 109 31 hrs 1 tera 10 1011 129 days 101 tera 12 1013 35 yrs 10 peta 14 1015 3523 yrs 1 exa Rely only on problem descriptionInformed Methods: Heuristic Search Idea: Informed search by using problem-specific knowledge. Best-First Search: Nodes are selected for expansion based on an evaluation function, f(n). Traditionally, f is a cost measure. Heuristic: Problem specific knowledge that (tries to) lead the search algorithm faster towards a goal state. Often implemented via heuristic function h(n). → Heuristic search is an attempt to search the most promising paths first. Uses heuristics, or rules of thumb, to find the best node to expand next.Uniform-cost search is NOT heuristic search s s s s 0 A A A A B B B C C C C G G G G 1 5 5 5 5 5 15 15 15 15 11 11 10 s 1 B 10 Requirement: g(Successor(n)) g(n) It only looks backwards; has no ability to predict future costs. Always expand lowest cost node in open-list. Goal-test only before expansion, not after generation.Generic Best-First Search 1. Set L to be the initial node(s) representing the initial state(s). 2. If L is empty, fail. Let n be the node on L that is ``most promising'‘ according to f. Remove n from L. 3. If n is a goal node, stop and return it (and the path from the initial node to n). 4. Otherwise, add successors(n) to L. Return to step 2. seemingly-best-first...A good heuristic • Heuristic cost should never overestimate the actual cost of a node – I.e. it must be “optimistic” – So that we never overlook a node that is actually goodGreedy Best-First Search Heuristic function h(n): estimated cost from node n to nearest goal node. Greedy Search: Let f(n) = h(n). Example: 8-puzzle 4 4 5 6 1 7 3 2 8 2 1 8 3 7 6 5 Start State Goal StateWhich move is better? Heuristic: # of incorrect tiles • A: Move space Left • B: Move space Down • C: Both the sameWhich move is better? Heuristic: Manhattan distance of tiles from correct location • A: Move space Left • B: Move space Down • C: Both the same1. hC = number of misplaced tiles 2. hM = Manhattan distance Are they admissible? Which one should we use? hC ≤ hM ≤ h* 8-PuzzleSearch Costs on 8-Puzzle h1: number of misplaced tiles h2: Manhattan distanceRoute Planning Greedy Best-first searchStraight line distances to BucharestRoute Planning Greedy Best-first search 380 374 366 329 193 160 242 253 176 374 100 244 241 77 0 80 Blue – Straight line distance to BucharestA* Search Idea: Use total estimated solution cost: g(n): Cost of reaching node n from initial node h(n): Estimated cost from node n to nearest goal A* evaluation function: f(n) = g(n) + h(n) → f(n) is estimated cost of cheapest solution through n.Admissibility h*(n) Actual cost to reach a goal from n. Definition: A heuristic function h is optimistic or admissible if h(n) ≤ h*(n) for all nodes n. (h never overestimates the cost of reaching the goal.) Theorem: If h is admissible, then the A* algorithm will never return a suboptimal goal node.Route Planning A* 380 374 366 329 193 160 242 253 176 374 100 244 241 77 0 80 Blue – Straight line distance to BucharestAssume: h admissible; f non-decreasing along any path. Proof: Assume C* is cost of optimal solution, G2 is suboptimal goal (so h(G2)=0) f(G2)=g(G2)+h(G2)=g(G2) > C* Assume node n is some node on the optimal path f(n)=g(n)+h(n) C* So f(n)  C* < f(G2) so n will always be expanded before G2. Proof of the optimality of A*A* Optimal: yes Complete: Unless there are infinitely many nodes with f(n)<f*. Assume locally finite: (1) finite branching, (2) every operator costs at least Complexity (time and space): Still exponential because of breadth-first nature. Unless |h(n) – h*(n)| ≤ O(log(h*(n))), with h* true cost of getting to goal. A* is optimally efficient: given the information in h, no other optimal search method can expand fewer nodes. 0Find the flat pattern that minimizes total welding length Example: Optimal flat pattern problemPermutation space Lots more…Searching the permutation space • Searching routes: Reduce travel distance. Heuristic estimates remaining min distance (optimistic: does not overestimate remaining distance) • Searching flat patterns: Reduce welding length. Heuristic estimates max savings (optimistic: does not underestimate remaining savings)Finding optimal unfoldingEffective Branching Factor N=1+b*+(b*)2+...+(b*)d = (1- bd+1)/(1-b) b* =~ N 1/d Branching factor of chess is about 35IDA* Memory is a problem for the A* algorithms. IDA* is like iterative deepening, but uses an f-cost limit rather than a depth limit. At each iteration, the cutoff value is the smallest f-cost of any node that exceeded the cutoff on the previous iteration. Each iteration uses conventional depth-first search.Example: IDA* • Initial state: A, f=100 • Cutoff: 100 • Cutoff: 101 • Cutoff: 105 A E D C B 100 101 110 200 105 A E D C B 100 101 110 200 105 G 120 H 105Similar to a DFS, but keeps track of the f-value of the best alternative path available from any ancestor of the current node. If current node exceeds this limit, recursion unwinds back to the alternative path, replacing the f-value of each node along the path with the best (highest, most accurate estimate) f-value of its children. (RBFS remembers the f-value of the best leaf in the forgotten subtree.) Recursive best-first search (RBFS)Example: RBFS A E D C B f=100 101 110 200 230 G 120 H 105 I 115 J 130 A E D C B f=100 115 110 200 230 K 117 L 140 L=[(C,110), (I,115), (G,120), (J,130), (D,200), (E,230)] L=[(C,110), (B,115), (D,200), (E,230)] L=[(B,115), (K,117), (L,140), (D,200), (E,230)] L=[(B,115), (C,117), (D,200), (E,230)]Example: RBFS (continued) A E D C B f=100 115 117 200 230 G 120 H 115 I 115 J 130 M Goal 116 N 120SMA* Simplified Memory-Bounded A* Search: • While memory available, proceeds just like A*, expanding the best leaf. • If memory is full, drops the worst leaf node - the one the highest f-cost; and stores this value in its parent node. (Won't know which way to go from this node, but we will have some idea of how worthwhile it is to explore the node.)Constructing Admissible Heuristics Deriving admissible heuristics automaticallyCombining Heuristics • h(n)=max(h1(n),h2(n),...,hm(n))Relaxed problems – A problem


View Full Document

CORNELL CS 4700 - Study Notes

Download Study Notes
Our administrator received your request to download this document. We will send you the file to your email shortly.
Loading Unlocking...
Login

Join to view Study Notes and access 3M+ class-specific study document.

or
We will never post anything without your permission.
Don't have an account?
Sign Up

Join to view Study Notes 2 2 and access 3M+ class-specific study document.

or

By creating an account you agree to our Privacy Policy and Terms Of Use

Already a member?