1. February 18, 2016Introduction to Artificial Intelligence Lecture 8: Search in State Spaces III 1 A General Backtracking Algorithm Sanity check function for the n-queens problem: fQueens :: Int -> [Int] -> Bool fQueens row plan = isOK plan 1 where isOK [] dist = True where isOK [] dist = True isOK (p:ps) dist = isOK (p:ps) dist = if elem p [row, row-dist, row+dist] then False if elem p [row, row-dist, row+dist] then False else isOK ps (dist + 1) else isOK ps (dist + 1)

2. February 18, 2016Introduction to Artificial Intelligence Lecture 8: Search in State Spaces III 2 A General Backtracking Algorithm We can write our general backtracking function: backtrack :: (a->[a]->Bool) -> [[a]] -> [a] backtrack fCheck levels = bt [] levels where bt plan [] = plan where bt plan [] = plan bt plan ([]:_) = [] bt plan ([]:_) = [] bt plan ((b:bs):ls) = if not (fCheck b plan) || null result bt plan ((b:bs):ls) = if not (fCheck b plan) || null result then bt plan (bs:ls) then bt plan (bs:ls) else result else result where result = bt (b:plan) ls where result = bt (b:plan) ls Use it to solve n-queens problem: nQueens n = backtrack fQueens $ take n (repeat [1..n])

3. February 18, 2016Introduction to Artificial Intelligence Lecture 8: Search in State Spaces III 3 A General Backtracking Algorithm Here is some code to print nQueens solution boards (rotated by 90  to simplify things): showQueens :: Int -> [Int] -> String showQueens size [] = "" showQueens size (q:qs) = (concat $ replicate (q-1) ". ") ++ "Q " ++ (concat $ replicate (size-q) ". ") ++ "\n" ++ (concat $ replicate (size-q) ". ") ++ "\n" ++ showQueens size qs showQueens size qs nQueensBoard :: Int -> String nQueensBoard n = showQueens n $ nQueens n The Haskell Platform allows you to even solve the 22-queens problem within about 15 seconds: main = putStrLn $ nQueensBoard 22

4. February 18, 2016Introduction to Artificial Intelligence Lecture 8: Search in State Spaces III 4 Our 22-Queens Solution.............. Q................. Q................. Q.............................. Q........... Q........................ Q........................ Q.............................. Q........ Q............................... Q......................... Q.................. Q............... Q............................ Q.................. Q.................. Q................. Q............... Q................... Q........................ Q................... Q................... Q.....................

5. February 18, 2016Introduction to Artificial Intelligence Lecture 8: Search in State Spaces III 5 A General Graph-Searching Algorithm We just saw how backtracking can lead to very efficient search when we continuously check whether we are stuck in a dead end. We would not be able to solve the 22-queens problem with brute force, i.e., breadth-first search. That would create a tree with at least 1 + 22 + 22 2 + … + 22 21 + 1 = (22 22 -1)/21 + 1  1.6  10 28 nodes before we could possibly find a solution. Even if we could create a billion nodes per second and store them somehow, it would still take us over 500 billion years to build the complete tree.

6. February 18, 2016Introduction to Artificial Intelligence Lecture 8: Search in State Spaces III 6 A General Graph-Searching Algorithm In some cases, we may even have an idea which decisions during search are most likely to get us to the goal quickly. Then we should take advantage of this information. Unfortunately, neither depth-first nor breadth-first search are flexible enough to allow this. Instead, we need a general graph-searching algorithm. It can be tuned towards depth-first or breadth-first search but also allows intermediate variants that can be guided by task- specific search heuristics. Let’s see how this magic works…

7. February 18, 2016Introduction to Artificial Intelligence Lecture 8: Search in State Spaces III 7 A General Graph-Searching Algorithm 1.Create search tree Tr consisting only of the start node (root) n 0. Put n 0 on an ordered list called OPEN. 2.Create empty list called CLOSED. 3.If OPEN is empty, exit with failure. 4.Move the first node on OPEN, called n, to CLOSED. 5.If n is a goal node, exit with success; the solution is the path from n to n 0 along the edges in Tr. 6.Expand node n by generating a set M of successors and connect them to n. Also put the members of M on OPEN. 7.Reorder the list OPEN according to a specific rule. 8.Go to step 3.

8. February 18, 2016Introduction to Artificial Intelligence Lecture 8: Search in State Spaces III 8 Rules for the Graph-Searching Algorithm What happens if we simply append new nodes to the end of OPEN? The algorithm turns into breadth-first search. The algorithm turns into breadth-first search. What happens if we add new nodes at the beginning of OPEN? The algorithm turns into depth-first search. The algorithm turns into depth-first search. What happens if we sort the nodes according to a heuristic evaluation function f’ ? The algorithm turns into best-first search. The algorithm turns into best-first search.

9. February 18, 2016Introduction to Artificial Intelligence Lecture 8: Search in State Spaces III 9 Best-First (Heuristic) Search We can often improve search efficiency by implementing a heuristic evaluation function f’ that estimates which node is the best one to expand in a given situation.We can often improve search efficiency by implementing a heuristic evaluation function f’ that estimates which node is the best one to expand in a given situation. Small values of f’ indicate the best nodes.Small values of f’ indicate the best nodes. In the general graph-searching algorithm, sort the nodes on the OPEN list in ascending order with regard to their f’-values.In the general graph-searching algorithm, sort the nodes on the OPEN list in ascending order with regard to their f’-values. This way, the supposedly “best” nodes are expanded first.This way, the supposedly “best” nodes are expanded first.

10. February 18, 2016Introduction to Artificial Intelligence Lecture 8: Search in State Spaces III 10 Best-First (Heuristic) Search Example: The Eight-puzzle The task is to transform the initial puzzle 283 164 75 into the goal state 1238 4 765 by shifting numbers up, down, left or right.

11. February 18, 2016Introduction to Artificial Intelligence Lecture 8: Search in State Spaces III 11 Best-First (Heuristic) Search How can we find an appropriate evaluation function? Heuristic evaluation functions have to be “custom- made” for individual problems.Heuristic evaluation functions have to be “custom- made” for individual problems. Often, it is easy to see what a reasonable (but usually not optimal) function could look like.Often, it is easy to see what a reasonable (but usually not optimal) function could look like. For the Eight-puzzle, it seems like a good idea to expand those nodes first that lead to configurations similar to the goal state.For the Eight-puzzle, it seems like a good idea to expand those nodes first that lead to configurations similar to the goal state. In our first attempt, we will simply define f’ as the number of digits in the puzzle that are not in the correct (goal) position yet.In our first attempt, we will simply define f’ as the number of digits in the puzzle that are not in the correct (goal) position yet.

12. February 18, 2016Introduction to Artificial Intelligence Lecture 8: Search in State Spaces III 12 Best-First (Heuristic) Search 283 164 75 4283164 752831 4 765283164 75 535283 14 7652 3184 76528314 765 3 to the goal 4 83214 765 3 3283714 65 48 3214 765 3 to further unnecessary expansions

13. February 18, 2016Introduction to Artificial Intelligence Lecture 8: Search in State Spaces III 13 Best-First (Heuristic) Search You see that with this simple function f’ the algorithm is still likely to perform unnecessarily many steps.You see that with this simple function f’ the algorithm is still likely to perform unnecessarily many steps. We should increase the “cost” of searches in greater depth to make the algorithm avoid them.We should increase the “cost” of searches in greater depth to make the algorithm avoid them. To achieve this, we simply add the current depth of a node to its f’-value.To achieve this, we simply add the current depth of a node to its f’-value. Now f’(n) = depth of n in the tree + number of digits in the puzzle in incorrect position.Now f’(n) = depth of n in the tree + number of digits in the puzzle in incorrect position. Let us try again!Let us try again!

14. February 18, 2016Introduction to Artificial Intelligence Lecture 8: Search in State Spaces III 14 Best-First (Heuristic) Search 283 164 75 0+4283164 752831 4 765283164 75 1+51+31+5283 14 7652 3184 76528314 765 2+3 goal ! 2+4 83214 765 3+3 2+3283714 65 3+4 23184 765 3+2123 84 765 4+11238 4 765 5+0123784 65 5+223184 765 3+4

15. February 18, 2016Introduction to Artificial Intelligence Lecture 8: Search in State Spaces III 15 Best-First (Heuristic) Search Using the improved function f’, the best-first algorithm discovers the solution much more efficiently than it did before.Using the improved function f’, the best-first algorithm discovers the solution much more efficiently than it did before. To formalize things, we now have a function f’(n) = g’(n) + h’(n), where g’(n) is a estimate of the depth (or cost) of n, and h’(n) is an estimate of the minimal distance (or cost) from n to a goal node.To formalize things, we now have a function f’(n) = g’(n) + h’(n), where g’(n) is a estimate of the depth (or cost) of n, and h’(n) is an estimate of the minimal distance (or cost) from n to a goal node. While f’(n), g’(n), and h’(n) are estimates, f(n) = g(n) + h(n) is the actual minimal cost of any path going from n 0 through n to a goal node.While f’(n), g’(n), and h’(n) are estimates, f(n) = g(n) + h(n) is the actual minimal cost of any path going from n 0 through n to a goal node.

16. February 18, 2016Introduction to Artificial Intelligence Lecture 8: Search in State Spaces III 16 Best-First (Heuristic) Search Question: What happens if for our heuristic estimation function f’(n) = g’(n) + h’(n) we simply set h’(n) = 0? Answer: Only the distance of a node from n 0 determines the cost of expanding a node. Only the distance of a node from n 0 determines the cost of expanding a node. Nodes closer to n 0 will be expanded first. Nodes closer to n 0 will be expanded first. The algorithm turns into breadth-first search. The algorithm turns into breadth-first search.

17. February 18, 2016Introduction to Artificial Intelligence Lecture 8: Search in State Spaces III 17 Best-First (Heuristic) Search And what happens if the state-space graph for our problem is not a tree? For example, is the Eight-puzzle state-space graph a tree? No. In this puzzle, actions are reversible. Any successor of a node n has n as a successor. In order to overcome this problem, we have to modify Step 6 in our general graph-searching algorithm.

18. February 18, 2016Introduction to Artificial Intelligence Lecture 8: Search in State Spaces III 18 Improved Graph-Searching Algorithm 1.Create search tree Tr consisting only of the start node (root) n 0. Put n 0 on an ordered list called OPEN. 2.Create empty list called CLOSED. 3.If OPEN is empty, exit with failure. 4.Move the first node on OPEN, called n, to CLOSED. 5.If n is a goal node, exit with success; the solution is the path from n to n 0 along the edges in Tr. 6.Expand node n by generating a set M of successors and connect them to n. Also put the members of M on OPEN. 7.Reorder the list OPEN according to a specific rule. 8.Go to step 3. 6.Expand node n by generating a set M of successors that are not already ancestors of n in Tr. Connect each node in M to node n. Also put the members of M on OPEN.

19. February 18, 2016Introduction to Artificial Intelligence Lecture 8: Search in State Spaces III 19 Algorithm A* The algorithm shown on the previous slide, using a heuristic estimation function f’(n) = g’(n) + h’(n), is also called Algorithm A*. Under the following conditions, A* is guaranteed to find the minimal cost path: Each node in the graph has a finite number of successors (or no successors). Each node in the graph has a finite number of successors (or no successors). All edges in the graph have costs greater than some positive amount . All edges in the graph have costs greater than some positive amount . For all nodes n in the search graph, h’(n)  h(n), that is, h’(n) never overestimates the actual value h(n) – an optimistic estimator. For all nodes n in the search graph, h’(n)  h(n), that is, h’(n) never overestimates the actual value h(n) – an optimistic estimator.

20. February 18, 2016Introduction to Artificial Intelligence Lecture 8: Search in State Spaces III 20 Algorithm A* Question: What happens if we use a “stupid” estimation function such as h’(n) = 0 ? Answer: A* is still guaranteed to find a minimal cost path. However, the more closely our function h’(n) is to the actual function h(n), the better will be the efficiency of A* in finding such a path. The effort of finding a good estimator pays off in the efficiency of the algorithm.