Artificial intelligence: informed search Outline Informed = use problem-specific knowledge Which search strategies? Best-first search and its variants Heuristic functions? How to invent them Local search and optimization Hill climbing, local beam search, genetic algorithms, Local search in continuous spaces Online search agents AI 1 1 maart 20

20 Pag.2 Previously: tree-search function TREE-SEARCH(problem,fringe) return a solution or failure fringe INSERT(MAKE-NODE(INITIAL-STATE[problem]), fringe) loop do if EMPTY?(fringe) then return failure node REMOVE-FIRST(fringe) if GOAL-TEST[problem] applied to STATE[node] succeeds then return SOLUTION(node) fringe INSERT-ALL(EXPAND(node, problem), fringe) A strategy is defined by picking the order of node expansion AI 1 1 maart 20

20 Pag.3 Best-first search General approach of informed search: Best-first search: node is selected for expansion based on an evaluation function f(n) Idea: evaluation function measures distance to the goal. Choose node which appears best Implementation: fringe is queue sorted in decreasing order of desirability. Special cases: greedy search, A* search AI 1

1 maart 20 20 Pag.4 A heuristic function [dictionary]A rule of thumb, simplification, or educated guess that reduces or limits the search for solutions in domains that are difficult and poorly understood. h(n) = estimated cost of the cheapest path from node n to goal node. If n is goal then h(n)=0 More information later. AI 1 1 maart 20 20

Pag.5 Romania with step costs in km hSLD=straight-line distance heuristic. hSLD can NOT be computed from the problem description itself In this example f(n)=h(n) Expand node that is closest to goal = Greedy best-first search AI 1 1 maart 20 20

Pag.6 Greedy search example Arad (366) Assume that we want to use greedy search to solve the problem of travelling from Arad to Bucharest. The initial state=Arad AI 1 1 maart 20 20 Pag.7 Greedy search example Arad

Zerind(374) Sibiu(253) Timisoara (329) The first expansion step produces: Sibiu, Timisoara and Zerind Greedy best-first will select Sibiu. AI 1 1 maart 20 20 Pag.8 Greedy search example

Arad Sibiu Arad (366) Fagaras (176) Oradea (380) Rimnicu Vilcea (193) If Sibiu is expanded we get: Arad, Fagaras, Oradea and Rimnicu Vilcea

Greedy best-first search will select: Fagaras AI 1 1 maart 20 20 Pag.9 Greedy search example Arad Sibiu Fagaras Sibiu (253) Bucharest (0)

If Fagaras is expanded we get: Sibiu and Bucharest Goal reached !! Yet not optimal (see Arad, Sibiu, Rimnicu Vilcea, Pitesti) AI 1 1 maart 20 20 Pag.10 Greedy search, evaluation Completeness: NO (cfr. DF-search) Check on repeated states Minimizing h(n) can result in false starts, e.g. Iasi to Fagaras. AI 1 1 maart 20

20 Pag.11 Greedy search, evaluation Completeness: NO (cfr. DF-search) Time complexity? m O(b ) Cfr. Worst-case DF-search (with m is maximum depth of search space) Good heuristic can give dramatic improvement. AI 1 1 maart 20

20 Pag.12 Greedy search, evaluation Completeness: NO (cfr. DF-search) Time complexity: O(b m ) m Space complexity: O(b ) Keeps all nodes in memory AI 1 1 maart 20 20

Pag.13 Greedy search, evaluation Completeness: NO (cfr. DF-search) m O(b ) Time complexity: O(b m ) Space complexity: Optimality? NO Same as DF-search AI 1 1 maart 20 20

Pag.14 A* search Best-known form of best-first search. Idea: avoid expanding paths that are already expensive. Evaluation function f(n)=g(n) + h(n) g(n) the cost (so far) to reach the node. h(n) estimated cost to get from the node to the goal. f(n) estimated total cost of path through n to goal. AI 1 1 maart 20 20 Pag.15

A* search A* search uses an admissible heuristic A heuristic is admissible if it never overestimates the cost to reach the goal Are optimistic Formally: 1. h(n) <= h*(n) where h*(n) is the true cost from n 2. h(n) >= 0 so h(G)=0 for any goal G. e.g. hSLD(n) never overestimates the actual road distance AI 1 1 maart 20 20 Pag.16 Romania example

AI 1 1 maart 20 20 Pag.17 A* search example Find Bucharest starting at Arad f(Arad) = c(??,Arad)+h(Arad)=0+366=366 AI 1 1 maart 20 20 Pag.18 A* search example

Expand Arrad and determine f(n) for each node f(Sibiu)=c(Arad,Sibiu)+h(Sibiu)=140+253=393 f(Timisoara)=c(Arad,Timisoara)+h(Timisoara)=118+329=447 f(Zerind)=c(Arad,Zerind)+h(Zerind)=75+374=449 Best choice is Sibiu AI 1 1 maart 20 20 Pag.19 A* search example Expand Sibiu and determine f(n) for each node

f(Arad)=c(Sibiu,Arad)+h(Arad)=280+366=646 f(Fagaras)=c(Sibiu,Fagaras)+h(Fagaras)=239+179=415 f(Oradea)=c(Sibiu,Oradea)+h(Oradea)=291+380=671 f(Rimnicu Vilcea)=c(Sibiu,Rimnicu Vilcea)+ h(Rimnicu Vilcea)=220+192=413 Best choice is Rimnicu Vilcea AI 1 1 maart 20 20 Pag.20 A* search example

Expand Rimnicu Vilcea and determine f(n) for each node f(Craiova)=c(Rimnicu Vilcea, Craiova)+h(Craiova)=360+160=526 f(Pitesti)=c(Rimnicu Vilcea, Pitesti)+h(Pitesti)=317+100=417 f(Sibiu)=c(Rimnicu Vilcea,Sibiu)+h(Sibiu)=300+253=553 Best choice is Fagaras AI 1 1 maart 20 20 Pag.21 A* search example Expand Fagaras and determine f(n) for each node f(Sibiu)=c(Fagaras, Sibiu)+h(Sibiu)=338+253=591

f(Bucharest)=c(Fagaras,Bucharest)+h(Bucharest)=450+0=450 Best choice is Pitesti !!! AI 1 1 maart 20 20 Pag.22 A* search example Expand Pitesti and determine f(n) for each node f(Bucharest)=c(Pitesti,Bucharest)+h(Bucharest)=418+0=418 Best choice is Bucharest !!! Optimal solution (only if h(n) is admissable)

Note values along optimal path !! AI 1 1 maart 20 20 Pag.23 Optimality of A*(standard proof) Suppose suboptimal goal G2 in the queue. Let n be an unexpanded node on a shortest to optimal goal G. f(G2 ) = g(G2 ) since h(G2 )=0 > g(G) since G2 is suboptimal >= f(n)since h is admissible Since f(G2) > f(n), A* will never select G2 for expansion

AI 1 1 maart 20 20 Pag.24 BUT graph search Discards new paths to repeated state. Previous proof breaks down Solution: Add extra bookkeeping i.e. remove more expsive of two paths. Ensure that optimal path to any repeated state is always first followed. Extra requirement on h(n): consistency (monotonicity) AI 1

1 maart 20 20 Pag.25 Consistency A heuristic is consistent if h(n) c(n,a,n') h(n') If h is consistent, we have f (n') g(n') h(n') g(n) c(n,a,n') h(n') g(n) h(n) f (n) i.e. f(n) is nondecreasing along any path.

AI 1 1 maart 20 20 Pag.26 Optimality of A*(more usefull) A* expands nodes in order of increasing f value Contours can be drawn in state space Uniform-cost search adds circles. F-contours are gradually Added: 1) nodes with f(n)

Nodes with f=fi, where fi < fi+1. AI 1 1 maart 20 20 Pag.27 A* search, evaluation Completeness: YES Since bands of increasing f are added Unless there are infinitly many nodes with f

Pag.28 A* search, evaluation Completeness: YES Time complexity: Number of nodes expanded is still exponential in the length of the solution. AI 1 1 maart 20 20 Pag.29 A* search, evaluation Completeness: YES Time complexity: (exponential with path length)

Space complexity: It keeps all generated nodes in memory Hence space is the major problem not time AI 1 1 maart 20 20 Pag.30 A* search, evaluation Completeness: YES Time complexity: (exponential with path length) Space complexity:(all nodes are stored) Optimality: YES

Cannot expand fi+1 until fi is finished. A* expands all nodes with f(n)< C* A* expands some nodes with f(n)=C* A* expands no nodes with f(n)>C* Also optimally efficient (not including ties) AI 1 1 maart 20 20 Pag.31 Memory-bounded heuristic search Some solutions to A* space problems (maintain completeness and optimality)

Iterative-deepening A* (IDA*) Here cutoff information is the f-cost (g+h) instead of depth Recursive best-first search(RBFS) Recursive algorithm that attempts to mimic standard best-first search with linear space. (simple) Memory-bounded A* ((S)MA*) Drop the worst-leaf node when memory is full AI 1 1 maart 20 20 Pag.32 Recursive best-first search function RECURSIVE-BEST-FIRST-SEARCH(problem) return a solution or failure

return RFBS(problem,MAKE-NODE(INITIAL-STATE[problem]),) function RFBS( problem, node, f_limit) return a solution or failure and a new f-cost limit if GOAL-TEST[problem](STATE[node]) then return node successors EXPAND(node, problem) if successors is empty then return failure, for each s in successors do f [s] max(g(s) + h(s), f [node]) repeat best the lowest f-value node in successors if f [best] > f_limit then return failure, f [best] alternative the second lowest f-value among successors result, f [best] RBFS(problem, best, min(f_limit, alternative)) if result failure then return result AI 1 1 maart 20 20

Pag.33 Recursive best-first search Keeps track of the f-value of the best-alternative path available. If current f-values exceeds this alternative fvalue than backtrack to alternative path. Upon backtracking change f-value to best fvalue of its children. Re-expansion of this result is thus still possible. AI 1 1 maart 20 20 Pag.34 Recursive best-first search, ex.

Path until Rumnicu Vilcea is already expanded Above node; f-limit for every recursive call is shown on top. Below node: f(n) The path is followed until Pitesti which has a f-value worse than the f-limit. AI 1 1 maart 20 20 Pag.35 Recursive best-first search, ex. Unwind recursion and store best f-value for current best leaf Pitesti result, f [best] RBFS(problem, best, min(f_limit, alternative)) best is now Fagaras. Call RBFS for new best

best value is now 450 AI 1 1 maart 20 20 Pag.36 Recursive best-first search, ex. Unwind recursion and store best f-value for current best leaf Fagaras result, f [best] RBFS(problem, best, min(f_limit, alternative)) best is now Rimnicu Viclea (again). Call RBFS for new best Subtree is again expanded.

Best alternative subtree is now through Timisoara. Solution is found since because 447 > 417. AI 1 1 maart 20 20 Pag.37 RBFS evaluation RBFS is a bit more efficient than IDA* Still excessive node generation (mind changes) Like A*, optimal if h(n) is admissible Space complexity is O(bd). IDA* retains only one single number (the current f-cost limit) Time complexity difficult to characterize

Depends on accuracy if h(n) and how often best path changes. IDA* en RBFS suffer from too little memory. AI 1 1 maart 20 20 Pag.38 (simplified) memory-bounded A* Use all available memory. I.e. expand best leafs until available memory is full When full, SMA* drops worst leaf node (highest f-value) Like RFBS backup forgotten node to its parent What if all leafs have the same f-value? Same node could be selected for expansion and deletion.

SMA* solves this by expanding newest best leaf and deleting oldest worst leaf. SMA* is complete if solution is reachable, optimal if optimal solution is reachable. AI 1 1 maart 20 20 Pag.39 Learning to search better All previous algorithms use fixed strategies. Agents can learn to improve their search by exploiting the meta-level state space. Each meta-level state is a internal (computational) state of a program that is searching in the object-level state space. In A* such a state consists of the current search tree

A meta-level learning algorithm from experiences at the meta-level. AI 1 1 maart 20 20 Pag.40 Heuristic functions E.g for the 8-puzzle Avg. solution cost is about 22 steps (branching factor +/- 3) Exhaustive search to depth 22: 3.1 x 1010 states. A good heuristic function can reduce the search process. AI 1

1 maart 20 20 Pag.41 Heuristic functions E.g for the 8-puzzle knows two commonly used heuristics h1 = the number of misplaced tiles h1(s)=8 h2 = the sum of the distances of the tiles from their goal positions (manhattan distance). h2(s)=3+1+2+2+2+3+3+2=18 AI 1 1 maart 20 20

Pag.42 Heuristic quality Effective branching factor b* Is the branching factor that a uniform tree of depth d would have in order to contain N+1 nodes. N 1 1 b * (b*) 2 ... (b*) d Measure is fairly constant for sufficiently hard problems. Can thus provide a good guide to the heuristics overall usefulness. A good value of b* is 1. AI 1 1 maart 20

20 Pag.43 Heuristic quality and dominance 1200 random problems with solution lengths from 2 to 24. If h2(n) >= h1(n) for all n (both admissible) then h2 dominates h1 and is better for search AI 1 1 maart 20 20 Pag.44 Inventing admissible heuristics Admissible heuristics can be derived from the exact

solution cost of a relaxed version of the problem: Relaxed 8-puzzle for h1 : a tile can move anywhere As a result, h1(n) gives the shortest solution Relaxed 8-puzzle for h2 : a tile can move to any adjacent square. As a result, h2(n) gives the shortest solution. The optimal solution cost of a relaxed problem is no greater than the optimal solution cost of the real problem. ABSolver found a usefull heuristic for the rubic cube. AI 1 1 maart 20 20 Pag.45 Inventing admissible heuristics

Admissible heuristics can also be derived from the solution cost of a subproblem of a given problem. This cost is a lower bound on the cost of the real problem. Pattern databases store the exact solution to for every possible subproblem instance. The complete heuristic is constructed using the patterns in the DB AI 1 1 maart 20 20 Pag.46 Inventing admissible heuristics Another way to find an admissible heuristic is through learning from experience: Experience = solving lots of 8-puzzles

An inductive learning algorithm can be used to predict costs for other states that arise during search. AI 1 1 maart 20 20 Pag.47 Local search and optimization Previously: systematic exploration of search space. Path to goal is solution to problem YET, for some problems path is irrelevant. E.g 8-queens Different algorithms can be used

Local search AI 1 1 maart 20 20 Pag.48 Local search and optimization Local search= use single current state and move to neighboring states. Advantages: Use very little memory Find often reasonable solutions in large or infinite state spaces. Are also useful for pure optimization problems.

Find best state according to some objective function. e.g. survival of the fittest as a metaphor for optimization. AI 1 1 maart 20 20 Pag.49 Local search and optimization AI 1 1 maart 20 20 Pag.50 Hill-climbing search

is a loop that continuously moves in the direction of increasing value It terminates when a peak is reached. Hill climbing does not look ahead of the immediate neighbors of the current state. Hill-climbing chooses randomly among the set of best successors, if there is more than one. Hill-climbing a.k.a. greedy local search AI 1 1 maart 20 20 Pag.51 Hill-climbing search function HILL-CLIMBING( problem) return a state that is a local

maximum input: problem, a problem local variables: current, a node. neighbor, a node. current MAKE-NODE(INITIAL-STATE[problem]) loop do neighbor a highest valued successor of current if VALUE [neighbor] VALUE[current] then return STATE[current] current neighbor AI 1 1 maart 20 20 Pag.52 Hill-climbing example

8-queens problem (complete-state formulation). Successor function: move a single queen to another square in the same column. Heuristic function h(n): the number of pairs of queens that are attacking each other (directly or indirectly). AI 1 1 maart 20 20 Pag.53 Hill-climbing example a) b)

a) shows a state of h=17 and the h-value for each possible successor. b) A local minimum in the 8-queens state space (h=1). AI 1 1 maart 20 20 Pag.54 Drawbacks Ridge = sequence of local maxima difficult for greedy algorithms to navigate Plateaux = an area of the state space where the evaluation function is flat. Gets stuck 86% of the time.

AI 1 1 maart 20 20 Pag.55 Hill-climbing variations Stochastic hill-climbing Random selection among the uphill moves. The selection probability can vary with the steepness of the uphill move. First-choice hill-climbing cfr. stochastic hill climbing by generating successors randomly until a better one is found. Random-restart hill-climbing Tries to avoid getting stuck in local maxima.

AI 1 1 maart 20 20 Pag.56 Simulated annealing Escape local maxima by allowing bad moves. Idea: but gradually decrease their size and frequency. Origin; metallurgical annealing Bouncing ball analogy: Shaking hard (= high temperature). Shaking less (= lower the temperature). If T decreases slowly enough, best state is reached. Applied for VLSI layout, airline scheduling, etc.

AI 1 1 maart 20 20 Pag.57 Simulated annealing function SIMULATED-ANNEALING( problem, schedule) return a solution state input: problem, a problem schedule, a mapping from time to temperature local variables: current, a node. next, a node. T, a temperature controlling the probability of downward steps current MAKE-NODE(INITIAL-STATE[problem]) for t 1 to do T schedule[t] if T = 0 then return current next a randomly selected successor of current

E VALUE[next] - VALUE[current] if E > 0 then current next else current next only with probability eE /T AI 1 1 maart 20 20 Pag.58 Local beam search Keep track of k states instead of one Initially: k random states

Next: determine all successors of k states If any of successors is goal finished Else select k best from successors and repeat. Major difference with random-restart search Information is shared among k search threads. Can suffer from lack of diversity. Stochastic variant: choose k successors at proportionally to state success. AI 1 1 maart 20 20 Pag.59 Genetic algorithms Variant of local beam search with sexual

recombination. AI 1 1 maart 20 20 Pag.60 Genetic algorithms Variant of local beam search with sexual recombination. AI 1 1 maart 20 20 Pag.61

Genetic algorithm function GENETIC_ALGORITHM( population, FITNESS-FN) return an individual input: population, a set of individuals FITNESS-FN, a function which determines the quality of the individual repeat new_population empty set loop for i from 1 to SIZE(population) do x RANDOM_SELECTION(population, FITNESS_FN) y RANDOM_SELECTION(population, FITNESS_FN) child REPRODUCE(x,y) if (small random probability) then child MUTATE(child ) add child to new_population population new_population until some individual is fit enough or enough time has elapsed return the best individual AI 1 1 maart 20

20 Pag.62 Exploration problems Until now all algorithms were offline. Offline= solution is determined before executing it. Online = interleaving computation and action Online search is necessary for dynamic and semi-dynamic environments It is impossible to take into account all possible contingencies. Used for exploration problems: Unknown states and actions. e.g. any robot in a new environment, a newborn baby,

AI 1 1 maart 20 20 Pag.63 Online search problems Agent knowledge: ACTION(s): list of allowed actions in state s C(s,a,s): step-cost function (! After s is determined) GOAL-TEST(s) An agent can recognize previous states. Actions are deterministic. Access to admissible heuristic h(s) e.g. manhattan distance AI 1

1 maart 20 20 Pag.64 Online search problems Objective: reach goal with minimal cost Cost = total cost of travelled path Competitive ratio=comparison of cost with cost of the solution path if search space is known. Can be infinite in case of the agent accidentally reaches dead ends AI 1 1 maart 20 20 Pag.65

The adversary argument Assume an adversary who can construct the state space while the agent explores it Visited states S and A. What next? Fails in one of the state spaces No algorithm can avoid dead ends in all state spaces. AI 1 1 maart 20 20 Pag.66 Online search agents The agent maintains a map of the environment.

Updated based on percept input. This map is used to decide next action. Note difference with e.g. A* An online version can only expand the node it is physically in (local order) AI 1 1 maart 20 20 Pag.67 Online DF-search function ONLINE_DFS-AGENT(s) return an action input: s, a percept identifying current state static: result, a table indexed by action and state, initially empty unexplored, a table that lists for each visited state, the action not yet tried unbacktracked, a table that lists for each visited state, the backtrack not yet tried

s,a, the previous state and action, initially null if GOAL-TEST(s) then return stop if s is a new state then unexplored[s] ACTIONS(s) if s is not null then do result[a,s] s add s to the front of unbackedtracked[s] if unexplored[s] is empty then if unbacktracked[s] is empty then return stop else a an action b such that result[b, s]=POP(unbacktracked[s]) else a POP(unexplored[s]) s s return a AI 1 1 maart 20 20 Pag.68

Online DF-search, example Assume maze problem on 3x3 grid. s = (1,1) is initial state Result, unexplored (UX), unbacktracked (UB), are empty S,a are also empty AI 1 1 maart 20 20 Pag.69 Online DF-search, example S=(1,1)

GOAL-TEST((,1,1))? S not = G thus false (1,1) a new state? True ACTION((1,1)) -> UX[(1,1)] {RIGHT,UP} s is null?

True (initially) UX[(1,1)] empty? False POP(UX[(1,1)])->a A=UP s = (1,1) Return a AI 1 1 maart 20 20

Pag.70 Online DF-search, example S=(2,1) GOAL-TEST((2,1))? S not = G thus false (2,1) a new state? True ACTION((2,1)) -> UX[(2,1)]

S {DOWN} s is null? false (s=(1,1)) result[UP,(1,1)] <- (2,1) UB[(2,1)]={(1,1)} UX[(2,1)] empty? False

A=DOWN, s=(2,1) return A AI 1 1 maart 20 20 Pag.71 Online DF-search, example S=(1,1) GOAL-TEST((1,1))? S not = G thus false (1,1) a new state?

false s is null? S AI 1 1 maart 20 20 Pag.72 false (s=(2,1)) result[DOWN,(2,1)] <- (1,1)

UB[(1,1)]={(2,1)} UX[(1,1)] empty? False A=RIGHT, s=(1,1) return A Online DF-search, example S=(1,2) GOAL-TEST((1,2))? S not = G thus false (1,2) a new state? True, UX[(1,2)]={RIGHT,UP,LEFT}

s is null? S false (s=(1,1)) result[RIGHT,(1,1)] <- (1,2) UB[(1,2)]={(1,1)} UX[(1,2)] empty? False A=LEFT, s=(1,2) return A AI 1 1 maart 20 20 Pag.73

Online DF-search, example S=(1,1) GOAL-TEST((1,1))? S not = G thus false (1,1) a new state? false s is null?

false (s=(1,2)) result[LEFT,(1,2)] <- (1,1) UB[(1,1)]={(1,2),(2,1)} UX[(1,1)] empty? S True UB[(1,1)] empty? False A= b for b in result[b, (1,1)]=(1,2)

B=RIGHT A=RIGHT, s=(1,1) AI 1 1 maart 20 20 Pag.74 Online DF-search Worst case each node is visited twice. An agent can go on a long walk even when it is close to the solution. An online iterative deepening approach solves this problem.

Online DF-search works only when actions are reversible. AI 1 1 maart 20 20 Pag.75 Online local search Hill-climbing is already online One state is stored. Bad performancd due to local maxima Random restarts impossible. Solution: Random walk introduces exploration (can produce exponentially many steps)

AI 1 1 maart 20 20 Pag.76 Online local search Solution 2: Add memory to hill climber Store current best estimate H(s) of cost to reach goal H(s) is initially the heuristic estimate h(s) Afterward updated with experience (see below) Learning real-time A* (LRTA*) AI 1 1 maart 20 20

Pag.77 Learning real-time A* function LRTA*-COST(s,a,s,H) return an cost estimate if s is undefined the return h(s) else return c(s,a,s) + H[s] function LRTA*-AGENT(s) return an action input: s, a percept identifying current state static: result, a table indexed by action and state, initially empty H, a table of cost estimates indexed by state, initially empty s,a, the previous state and action, initially null if GOAL-TEST(s) then return stop if s is a new state (not in H) then H[s] h(s) unless s is null result[a,s] s H[s] MIN LRTA*-COST(s,b,result[b,s],H)

b ACTIONS(s) a an action b in ACTIONS(s) that minimizes LRTA*-COST(s,b,result[b,s],H) s s return a AI 1 1 maart 20 20 Pag.78