Today’s Class Local Search • Iterative improvement methods AI Class 6 (Ch. 4.1-4.2) • Hill climbing “If the path to the goal • Simulated annealing does not matter… [we • Local beam search can use] a single current node and move to • Genetic algorithms neighbors of that node.” • Online search – R&N pg. 121 Based on slides by Dr. Marie desJardin. Some material also adapted from slides by Dr. Matuszek @ Villanova University, which are based on Hwee Tou Ng at Berkeley, which Cynthia Matuszek – CMSC 671 are based on Russell at Berkeley. Some diagrams are based on AIMA. 3 Admissibility Admissibility and Optimality • Admissibility is a property of heuristics • Intuitively: • They are optimistic – think goal is closer than it is • When A* finds a path of length k , it has already tried • (Or, exactly right) every other path which can have length ≤ k • Because all frontier nodes have been sorted in ascending • Admissible algorithms order of f ( n )= g ( n )+ h ( n ) can be pretty bad! • Does an admissible heuristic guarantee optimality • Is h ( n ): “1 kilometer” admissible? for greedy search? • Using admissible heuristics guarantees that the first • Reminder: f ( n ) = h ( n ), always choose node “nearest” goal solution found will be optimal, for some algorithms (A*). • No sorting beyond that 4 5 Local Search Algorithms Local Search Algorithms Very efficient! • Sometimes the path to the goal is irrelevant • Sometimes the path to the goal is irrelevant • Goal state itself is the solution • Goal state itself is the solution Why? • an objective function to evaluate states E • an objective function to evaluate states E • In such cases, we can use local search algorithms • State space = set of “complete” configurations • That is, all elements of a solution are present • Keep a single “current” state, try to improve it • Find configuration satisfying constraints X • Example? • In such cases, we can use local search algorithms • Keep a single “current” state, try to improve it 6 7 1
State Space (Landscape) State Space (Landscape) S 2 A 1 B 4 B S A 8 Iterative Improvement Search Hill Climbing on State Surface Concept: • Start with an initial guess • trying to reach the “highest” • Gradually improve it until it is legal or optimal (most desirable) • Some examples: point (state) • Hill climbing • “Height” • Simulated annealing Defined by Evaluation • Constraint satisfaction Function 10 11 Hill Climbing Example Hill Climbing Search 2 8 3 1 2 3 start 1 6 4 8 4 h = 0 goal h = -4 • If there exists a successor s for the current state n such that 7 5 7 6 5 • h ( s ) > h ( n ) • h ( s ) >= h ( t ) for all the successors t of n , -2 -5 -5 then move from n to s . Otherwise, halt at n . 2 8 3 2 1 3 • Look one step ahead to determine if any successor is “better” than current state 1 4 h = -3 8 4 h = -1 • If so, move to the best successor 7 6 5 7 6 5 A kind of Greedy search in that it uses h • • But, does not allow backtracking or jumping to an alternative path -3 -4 • Doesn’t “remember” where it has been. 2 3 2 3 Not complete • 1 8 4 1 8 4 h = -2 • Search will terminate at local minima, plateaux, ridges. 7 6 5 7 6 5 h = -3 -4 12 f(n) = -(number of tiles out of place) 2
Exploring the Landscape Drawbacks of Hill Climbing local maximum • Local Maxima : • Problems: local maxima, plateaus, ridges • Peaks that aren’t the highest point in the space • Remedies: plateau • Plateaus: • Random restart: keep restarting the search from random • A broad flat region that gives locations until a goal is found. the search algorithm no direction (random walk) • Problem reformulation: reformulate the search space to eliminate these problematic features • Ridges: • Flat like a plateau, but with • Some problem spaces are great for hill climbing; drop-offs to the sides; steps ridge to the North, East, South others are terrible and West may go down, but a step to the NW may go up. Image from: http://classes.yale.edu/fractals/CA/GA/Fitness/Fitness.html 15 Example of a Local Optimum Some Extensions of Hill Climbing • Simulated Annealing 1 2 5 • Escape local maxima by allowing some “bad” moves but f = -7 7 4 gradually decreasing their frequency move start goal up 8 6 3 Local Beam Search • 1 2 5 1 2 3 • Keep track of k states rather than just one 8 7 4 8 4 f = 0 • At each iteration: move 6 3 7 6 5 right • All successors of the k states are generated and evaluated f = -6 1 2 5 • Best k are chosen for the next iteration 8 7 4 f = -7 6 3 f = -(manhattan distance) 16 17 Some Extensions of Hill Climbing Gradient Ascent / Descent • Take downward “steps” proportional to the negative of the • Stochastic Beam Search gradient of the function at current state. • Chooses semi-randomly from “uphill” possibilities • “Steepest descent” • “Steeper” moves have a higher probability of being chosen • Gradient descent procedure for finding the arg x min f(x) • Random-Restart Climbing • choose initial x 0 randomly • Can actually be applied to any form of search • repeat • Pick random starting points until one leads to a solution • x i+1 � x i – � f’ (x i ) • until the sequence x 0 , x 1 , …, x i , x i+1 converges • Genetic Algorithms • Step size � (eta) is • Each successor is generated from two predecessor (parent) small (~0.1–0.05) states • Good for differentiable, continuous spaces 18 19 Images from http://en.wikipedia.org/wiki/Gradient_descent 3
Gradient Ascent / Descent Gradient Methods vs. Newton’s Method • A reminder of Newton’s method from Calculus: x i+1 � x i – � f’ (x i ) / f’’ (x i ) • Newton � s method uses 2 nd order information (the second derivative, or, curvature ) to take a more direct route to the minimum. Contour lines of a function (blue) • The second-order information • Gradient descent (green) is more expensive to compute, • Newton’s method (red) but converges more quickly. 20 Images from http://en.wikipedia.org/wiki/Gradient_descent Images from http://en.wikipedia.org/wiki/Newton's_method_in_optimization Simulated Annealing Simulated Annealing (II) • Simulated annealing (SA): analogy between the way • Can avoid becoming trapped at local minima. metal cools into a minimum-energy crystalline structure • Uses a random local search that: and the search for a minimum generally • Accepts changes that increase objective function f • In very hot metal, molecules can move fairly freely • As well as some that decrease it • But, they are slightly less likely to move out of a stable structure • As you slowly cool the metal, more molecules are “trapped” in • Uses a control parameter T freedom to place • By analogy with the original application make “bad” • Conceptually: Escape local maxima by allowing some • Is known as the system “ temperature ” moves “bad” (locally counterproductive) moves but gradually decreasing their frequency • T starts out high and gradually decreases toward 0 22 23 Simulated Annealing (IV) Local Beam Search • f ( s ) represents the quality of state n (high is good) • Begin with k random states • A “bad” move from A to B is accepted with a probability • k , instead of one, current state(s) P(move A � B ) ≈ e ( f (B) – f (A)) / T • Generate all successors of these states • (Note that f (B) – f (A) will be negative, so bad moves always have a relative probability less than one. Good moves, for which f (B) – f (A) is positive, have a • Keep the k best states relative probability greater than one.) • Temperature • Stochastic beam search • Higher temperature = more likely to make a “bad” move • Probability of keeping a state is a function of its heuristic • As T tends to zero, this probability tends to zero value • SA becomes more like hill climbing • If T is lowered slowly enough, SA is complete and admissible. • More likely to keep “better” successors • domain-specific � • sometimes hard to determine 27 4
Recommend
More recommend
