1. 8:38:34 PMICS 573: High-Performance Computing11.1 Search Algorithms for Discrete Optimization Problems Discrete Optimization - Basics Sequential Search Algorithms Parallel Depth-First Search Parallel Best-First Search Speedup Anomalies in Parallel Search Algorithms

2. 8:38:34 PMICS 573: High-Performance Computing11.2 Discrete Optimization Problems Discrete optimization problems (DOPs) are a class of computationally expensive problems –of significant theoretical and practical interest. Characterized by a systematic search for solution from among many potential solutions, subject to constraints. For many such problems, exhaustive search is infeasible –Some form of directed search is performed, instead Examples from classical computer science problems –The traveling salesman problem –The 0/1 knapsack problem –The n-queens problem –The 15- and 8-puzzles Examples in commerce, banking and industry –Financial forecasting –Airline fleet and crew assignment –VLSI chip layout

3. 8:38:34 PMICS 573: High-Performance Computing11.3 Solution Techniques for DOPs Several search and optimization techniques can be used on DOPs: 1.Branch-and-bound (B&B) search: –may be described in terms of finding the minimal value of a function f(x) over a set of admissible values of the argument x called feasible region –A B&B procedure requires branching and bounding: Branching: a smart way of covering the feasible region by several smaller feasible subregions (ideally, splitting into subregions). Bounding: a fast way of finding upper and lower bounds for the optimal solution within a feasible region Recursive branching naturally results in a tree of subregions called a state space tree. The state space tree may need to be pruned for increased efficiency. 2.Dynamic programming: –Uses a top-down or bottom-up approach to reduce the runtime of algorithms having the following properties: Overlapping subproblems: the problem can be broken down into subproblems which are reused several times (e.g., Fibonacci). Optimal substructure: its optimal solution can be constructed efficiently from optimal solutions to its subproblems (e.g., shortest paths). Memoization may be used in overlapping subproblems for increased efficiency.

4. 8:38:34 PMICS 573: High-Performance Computing11.4 Solution Techniques for DOPs 3.Hill climbing –A graph search algorithm where the current path is extended with a successor node which is closer to the solution (without retaining its siblings or parent) –Used widely in AI for reaching a goal state from a starting node. –Limitation: may find only local maxima, failing to reach goal, global maxima. 4.Simulated annealing (SA) –A generic probabilistic algorithm for locating a good approximation to the global optimum of a given function in a large search space –Can be used to overcome the local maxima problem in hill climbing 5.Genetic algorithms (GAs) –A search technique used to find approximate solutions to DOPs –GAs use techniques inpired by evolutionary biology such as inheritance, mutation, natural selection, and recombination (or crossover) –Candidate solutions (called individuals) to an optimization problem evolve toward better solutions using fitness analyses.

5. 8:38:34 PMICS 573: High-Performance Computing11.5 Definitions A discrete optimization problem (DOP) is defined as a tuple of (S, f) –S : The set of feasible states – f : A cost function f : S -> R The objective of a DOP is to find a feasible solution x opt, such that f( x opt ) ≤ f( x ) for all x  S. Examples 0/1 Linear integer programming problem: Given an mxn matrix A, vectors b and c, find vector x such that x contains only 0s and 1s Ax > b f(x) = x T c is minimized. 8-puzzle problem: Given an initial configuration of an 8-puzzle find the shortest sequence of moves that will lead to the final configuration.

6. 8:38:34 PMICS 573: High-Performance Computing11.6 Discrete Optimization: Example

7. 8:38:34 PMICS 573: High-Performance Computing11.7 Formulating DOPs as Graph Search The feasible space S is typically very large. Many DOP can be formulated as finding the minimum cost path in a graph. –Nodes in the graph correspond to states. –States are classified as either terminal or non-terminal –Some of the states correspond to feasible solutions whereas others do not. –Edges correspond to “costs” associated with moving from one state to the other. These graphs are called state-space graphs.

8. 8:38:34 PMICS 573: High-Performance Computing11.8 Examples of State-Space Graphs

9. 8:38:34 PMICS 573: High-Performance Computing11.9 Examples of State-Space Graphs 0/1 integer linear programming problem –States correspond to partial assignment of values to components of the x vector.

10. 8:38:34 PMICS 573: High-Performance Computing11.10 Admissible Heuristics Often, it is possible to estimate the cost to reach the goal state from an intermediate state. This estimate, called a heuristic estimate, can be effective in guiding search to the solution. If the estimate is guaranteed to be an underestimate, the heuristic is called an admissible heuristic. Admissible heuristics have desirable properties in terms of optimality of solution (as we shall see later). An admissible heuristic for 8-puzzle is as follows: –Assume that each position in the 8-puzzle grid is represented as a pair. –The distance between positions (i,j) and (k,l) is defined as |i - k| + |j - l|. This distance is called the Manhattan distance. –The sum of the Manhattan distances between the initial and final positions of all tiles is an admissible heuristic.

11. 8:38:34 PMICS 573: High-Performance Computing11.11 Parallel Discrete Optimization: Motivation DOPs are generally NP-hard problems. Does parallelism really help much? –Cannot reduce their worst-case running time to a polynomial one For many problems, the average-case runtime is polynomial. Often, we can find suboptimal solutions in polynomial time. –Bigger problem instances can be solved Many DOPs have smaller state spaces but require real-time solutions. –Parallel processing may be the only way to obtain acceptable performance!

12. 8:38:34 PMICS 573: High-Performance Computing11.12 Exploring the State-Space Graphs The solution is discovered by exploring the state-space search. –Exponentially large Heuristic estimates of the solution cost are used. –Cost of reaching to a feasible solution from current state x is l(x) = g(x) + h(x) Admissible heuristics are the heuristics that correspond to lower bounds on the actual cost. –Manhattan distance is an admissible heuristic for the 8-puzzle problem. Idea is to explore the state-space graph using heuristic cost estimates to guide the search. –Do not spend any time exploring “bad” or “unpromising” states.

13. 8:38:34 PMICS 573: High-Performance Computing11.13 Search Space Structure: Trees Vs Graphs Is the search space a tree or a graph? Trees –Each new successor leads to an unexplored part of the search space. –Example: the space of a 0/1 integer linear programming Graphs –A state can be reached along multiple paths. –Whenever a state is generated, it is necessary to check if the state has already been generated. –Example: the space of 8-puzzle. Next, we consider sequential solution strategies for DOPs formulated as tree or graph search problems.

14. 8:38:34 PMICS 573: High-Performance Computing11.14 Sequential Exploration Strategies 1.Depth-First i.Simple Backtracking –Performs DFS until it finds the first feasible solution and terminates –Uses no heuristic to order the successors of the expanded node –Not guaranteed to find optimal solution –Ordered backtracking uses heuristics to order successors ii.Depth-First Branch-and-Bound (DFBB) –Partial solutions that are inferior to the current best solutions are discarded. –Does not explore paths leading to inferior solutions iii.Iterative Deepening –Tree is expanded up to certain depth. –If no feasible solution is found, the depth is increased and the entire process is repeated. –Useful when solution exists close to root on alternative branch iv.Iterative Deepening A* (IDA*) –Imposes cost-bound rather than a depth-bound –Finds an optimal solution when an admissible heuristic is used

15. 8:38:34 PMICS 573: High-Performance Computing11.15 DFS Storage Requirements and Data Structures Suitable primarily for state-graphs that are trees At each step of DFS, untried alternatives must be stored for backtracking. If m is the amount of storage required to store a state, and d is the maximum depth, then the total space requirement of the DFS algorithm is O(md). The state-space tree searched by parallel DFS can be efficiently represented as a stack. Memory requirement of the stack is linear in depth of tree.

16. 8:38:34 PMICS 573: High-Performance Computing11.16 Depth-First Search On 8-Puzzle

17. 8:38:34 PMICS 573: High-Performance Computing11.17 Representing a DFS Tree

18. 8:38:34 PMICS 573: High-Performance Computing11.18 Sequential Exploration Strategies 2.Best-First Search (BFS) –BFS algorithms use a heuristic to guide search. –If the heuristic is admissible, the BFS finds the optimal solution. –OPEN/CLOSED lists –A* algorithm Heuristic estimate is used to order the nodes in the open list. Suitable for state-space graphs that are either trees or graphs. Large memory complexity. –Proportional to the number of states visited.

19. 8:38:34 PMICS 573: High-Performance Computing11.19 Data Structures for BFS Algorithms The core data structure is a list, called open list, that stores unexplored nodes sorted on their heuristic estimates. The best node is selected from the list, expanded, and its off-spring are inserted at the right position. BFS of graphs must be slightly modified to account for multiple paths to the same node. A closed list stores all the nodes that have been previously seen. If a newly expanded node exists in the open or closed lists with better heuristic value, the node is not inserted into the open list.

20. 8:38:34 PMICS 573: High-Performance Computing11.20 The A* Algorithm A BFS technique that uses admissible heuristics. Defines function l(x) for each node x as g(x) + h(x). Here, g(x) is the cost of getting to node x and h(x) is an admissible heuristic estimate of getting from node x to the solution. The open list is sorted on l(x). The space requirement of BFS is exponential in depth!

21. 8:38:34 PMICS 573: High-Performance Computing11.21 Best-First Search On 8-Puzzle

22. 8:38:34 PMICS 573: High-Performance Computing11.22 Search Overhead Factor The amount of work done by serial and parallel formulations of search algorithms is often different. –Why? Let W be serial work and W P be parallel work. –Search overhead factor s is defined as W P /W. Upper bound on speedup is p × (W/W P ). Assumptions for subsequent analyses: –the time to expand each node, t c, is the same –W and W p are the number of nodes expanded by the serial and the parallel formulations, respectively. –From the above, the sequential run time is given by T S = t c W. –We assume that t c = 1.

23. 8:38:34 PMICS 573: High-Performance Computing11.23 Parallel Depth-First Search Challenges Computation is dynamic and unstructured –Why dynamic? –Why unstructured? Decomposition approaches? –Do we do the same work as the sequential algorithm? Mapping approaches? –How do we ensure load balance? –Static partitioning of unstructured trees yields poor performance. Dynamic load balancing is required.

24. 8:38:34 PMICS 573: High-Performance Computing11.24 Tree Search Can Be Unstructured

25. 8:38:34 PMICS 573: High-Performance Computing11.25 Parallel DFS: Dynamic Load Balancing When a processor runs out of work, it gets more work from another processor. –Trades communication cost for load balancing This is done using –work requests and responses in message passing machines, and –locking and extracting work in shared address space machines. On reaching final state at a processor, all processors terminate. Unexplored states can be conveniently stored as local stacks at processors. The entire space is assigned to one processor to begin with.

26. 8:38:34 PMICS 573: High-Performance Computing11.26 Parallel Depth-First Search: Dynamic Load Balancing

27. 8:38:34 PMICS 573: High-Performance Computing11.27 Performance of Parallel DFS Two important issues determine the performance of parallel DFS: 1.Work splitting strategies Which states from my stack should I give away? 2.Load balancing strategies Which processor should I ask for work? Work splitting strategies –Work is split by splitting the stack into two. –Ideally, we do not want either of the split pieces to be small. –Select nodes near the bottom of the stack (node splitting), or –Select some nodes from each level (stack splitting). –The second strategy generally yields a more even split of the space.

28. 8:38:34 PMICS 573: High-Performance Computing11.28 Parameters in Parallel DFS: Work Splitting

29. 8:38:34 PMICS 573: High-Performance Computing11.29 Load-Balancing Schemes Who do you request work from? –Note: we would like to distribute work requests evenly, in a global sense. Asynchronous round robin: –Each processor maintains a counter and makes requests in a round- robin fashion. –Work requests are generated independently by each processor. Global round robin: –The system maintains a global counter and requests are made in a round-robin fashion, globally. –Ensures that successive work requests are distributed evenly over all processors. Random polling: –Request a randomly selected processor for work.

30. 8:38:34 PMICS 573: High-Performance Computing11.30 Termination Detection How do you know when everyone's done? A number of algorithms have been proposed.

31. 8:38:34 PMICS 573: High-Performance Computing11.31 Dijkstra's Token Termination Detection Assume that all processors are organized in a logical ring. Assume, for now that work transfers can only happen from P i to P j if j > i. Processor P 0 initiates a token on the ring when it goes idle. Each intermediate processor receives this token and forwards it when it becomes idle. When the token reaches processor P 0, all processors are done.

32. 8:38:34 PMICS 573: High-Performance Computing11.32 Dijkstra's Token Termination Detection

33. 8:38:34 PMICS 573: High-Performance Computing11.33 Dijkstra's Token Termination Detection When processor P 0 goes idle, it colors itself green and initiates a green token. If processor P j sends work to processor P i and j > i then processor P j becomes red. If processor P i has the token and P i is idle, it passes the token to P i+1. –If P i is red, then the color of the token is set to red before it is sent to P i+1. –If P i is green, the token is passed unchanged. After P i passes the token to P i+1, P i becomes green. The algorithm terminates when processor P 0 receives a green token and is itself idle.

34. 8:38:34 PMICS 573: High-Performance Computing11.34 Dijkstra's Token Termination Detection

35. 8:38:34 PMICS 573: High-Performance Computing11.35 Tree-Based Termination Detection Associate weights with individual workpieces. Initially, processor P 0 has all the work and a weight of one. Whenever work is partitioned, the weight is split into half and sent with the work. When a processor gets done with its work, it sends its parent the weight back. Termination is signaled when the weight at processor P 0 becomes 1 again. Note that underflow and finite precision are important factors associated with this scheme.

36. 8:38:34 PMICS 573: High-Performance Computing11.36 Tree-Based Termination Detection

37. 8:38:34 PMICS 573: High-Performance Computing11.37 Parallel Formulations of Depth-First Branch-and- Bound Parallel formulations of depth-first branch-and-bound search (DFBB) are similar to those of DFS. Each processor has a copy of the current best solution. This is used as a local bound. If a processor detects another solution, it compares the cost with current best solution. –If the cost is better, it broadcasts this cost to all processors. If a processor's current best solution path is worse than the globally best solution path, only the efficiency of the search is affected, not its correctness.

38. 8:38:34 PMICS 573: High-Performance Computing11.38 Parallel Formulations of IDA* Common Cost Bound: –Each processor is given the same cost bound. –Processors use parallel DFS on the tree within the cost bound. –The drawback of this scheme is that there might not be enough concurrency. Variable Cost Bound: –Each processor works on a different cost bound. –The major drawback here is that a solution is not guaranteed to be optimal until all lower cost bounds have been exhausted. In each case, parallel DFS is the search kernel.

39. 8:38:34 PMICS 573: High-Performance Computing11.39 Parallel Best-First Search The core data structure is the open list (typically implemented as a priority queue). Each processor locks this queue, extracts the best node, unlocks it. Successors of the node are generated, their heuristic functions estimated, and the nodes inserted into the open list as necessary after appropriate locking. Since we expand more than one node at a time, we may expand nodes that would not be expanded by a sequential algorithm.

40. 8:38:34 PMICS 573: High-Performance Computing11.40 Parallel BFS: Centralized Lists

41. 8:38:34 PMICS 573: High-Performance Computing11.41 Problems with Centralized Lists The termination criterion of sequential BFS fails for parallel BFS. –p nodes from the open list are being expanded The open list is a point of contention. –Can severely limit speedup even on shared-address-space architectures Distributed open lists may address can help

42. 8:38:34 PMICS 573: High-Performance Computing11.42 Parallel Best-First Search Avoid contention by having multiple open lists. Initially, the search space is statically divided across these open lists. Processors concurrently operate on these open lists. Since the heuristic values of nodes in these lists may diverge significantly, –we must periodically balance the quality of nodes in each list. A number of balancing strategies based on ring, blackboard, or random communications are possible.

43. 8:38:34 PMICS 573: High-Performance Computing11.43 Parallel Best-First Search

44. 8:38:34 PMICS 573: High-Performance Computing11.44 Parallel Best-First Search

45. 8:38:34 PMICS 573: High-Performance Computing11.45 Parallel Best-First Graph Search Graph search involves a closed list, where the major operation is a lookup (on a key corresponding to the state). The classic data structure is a hash. Hashing can be parallelized by using two functions: –the first one hashes each node to a processor, and –the second one hashes within the processor. This strategy can be combined with the idea of multiple open lists. If a node does not exist in a closed list, it is inserted into the open list at the target of the first hash function. In addition to facilitating lookup, randomization also equalizes quality of nodes in various open lists.

46. 8:38:34 PMICS 573: High-Performance Computing11.46 Speedup Anomalies in Parallel Search Since the search space explored by processors is determined dynamically at runtime, the actual work might vary significantly. Speedup anomalies using p processors: –acceleration anomalies: executions yielding speedups greater than p –deceleration anomalies: executions yielding speedups of less than p Speedup anomalies also manifest themselves in best-first search algorithms. If the heuristic function is good, the work done in parallel best-first search is typically more than that in its serial counterpart.

47. 8:38:34 PMICS 573: High-Performance Computing11.47 Speedup Anomalies in Parallel Search

48. 8:38:34 PMICS 573: High-Performance Computing11.48 Speedup Anomalies in Parallel Search