iterative budgeted exponential search
play

Iterative Budgeted Exponential Search Malte Helmert 1 , Tor Lattimore - PowerPoint PPT Presentation

Mar 05, 2023 •461 likes •1.18k views

Iterative Budgeted Exponential Search Malte Helmert 1 , Tor Lattimore 2 , Levi H.S. Lelis 3 , Laurent Orseau 2 , Nathan Sturtevant 4 1 University of Basel, Switzerland 2 DeepMind, UK 3 Federal de Vi cosa, Brazil 4 University of Alberta, Canada

  1. Iterative Budgeted Exponential Search Malte Helmert 1 , Tor Lattimore 2 , Levi H.S. Lelis 3 , Laurent Orseau 2 , Nathan Sturtevant 4 1 University of Basel, Switzerland 2 DeepMind, UK 3 Federal de Vi¸ cosa, Brazil 4 University of Alberta, Canada IJCAI 2019 - Macau, China Talk jointly presented by Laurent Orseau and Nathan Sturtevant

  2. Start 5 4 3 2 1 h = 11 1 2 3 4 5 Goal

  3. Shortest Path Algorithm History 1960 1970 1980 1990 2000 2010 IDA* worst case IDA* IBEX A* A* worst case † Dijkstra’s Algorithm

  4. The N 2 problem N 2 • Algorithms perform � re-expansions of N states N 2 N = b d • IDA*: � expansions of � states 2 N • A*: � expansions of N states N 2 • B & B’: � (Martelli 1977; Mero, 1984) N 2 • Weighted A*: � (Chen et al, 2019) Iterative Budgeted Exponential Search 4

  5. IDA* Refresher • IDA* does iterative deepening search on f -costs • f(n) = g(n) + h(n) • Next iteration f -cost: • Smallest unexplored from previous iteration • If f -cost layers grow exponentially • IDA* is asymptotically optimal in node expansions Iterative Budgeted Exponential Search 5

  6. IDA* - Unit Costs

  7. f-cost 11 2 States x8 f-cost 13 16 States x5 79 States f-cost 15

  8. IDA* Worst Case • f -cost layers grow exponentially 1 + b + b 2 + b 3 + … + b d ≈ b d • � • What if f -cost layers grew linearly? 1 + 2 + 3 + 4 + … + b d ≈ ( b d ) 2 • � • Happens with non-unit edge costs: t + 2 • STP: Cost of moving tile t : � t + 1 Iterative Budgeted Exponential Search 8

  10. f-cost 11 11.25 13.5 13.7 13.45 13.83 13.87 13.62

  11. Previous Work • Iterative Deepening • IDA* (Korf, 1985) • IDA* CR (Sarkar et al, 1990) • IDA* IM (Burns & Ruml, 2013) • EDA* (Sharon et al, 2014) Iterative Budgeted Exponential Search 11

  13. A* (B/B’) • A* with a consistent heuristic is optimal* • With an inconsistent heuristic A* might do 2 N expansions of N states • B and B’ immediately propagate shorter paths • Will do at most N 2 expansions of N states Iterative Budgeted Exponential Search 13

  14. 0 N states 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 N states

  16. Previous Work • Inconsistent Heuristics • B/B’ (Martelli 1977; Mero, 1984) • Delay (Sturtevant et al, 2008) • BPMX (Felner et al, 2011) Iterative Budgeted Exponential Search 16

  18. Demos • All demos shown here runnable online: • https://www.movingai.com/SAS/BTS/ • Also running on my iPad; can give detailed demo Iterative Budgeted Exponential Search 18

  19. IDA* — Good Case

  20. IDA* — Good Case

  21. IDA* — Good Case

  22. IDA* — Good Case

  23. IDA* — Good Case

  24. IDA* — Good Case

  25. IDA* — Bad Case

  26. IDA* — Bad Case

  27. IDA* — Bad Case

  28. IDA* — Bad Case

  29. IDA* — Bad Case

  30. IDA* — Bad Case

  31. IDA* — Bad Case

  32. IDA* — Bad Case

  33. IDA* — Bad Case

  34. EDA*

  35. EDA*

  36. EDA*

  37. EDA*

  38. EDA*

  39. EDA*

  40. EDA*

  41. IBEX (main idea) For budget = 2, 4, 8, 16, ...: cost_bound = Oracle(budget) search within cost_bound and within budget Oracle(budget) returns largest cost for which budget is su ffi cient.

  42. IBEX (main idea): With an oracle budget f -cost expansions 2

  43. IBEX (main idea): With an oracle budget f -cost expansions 2 11.2

  44. IBEX (main idea): With an oracle budget f -cost expansions 2 11.2 2

  45. IBEX (main idea): With an oracle budget f -cost expansions 2 11.2 2 4

  46. IBEX (main idea): With an oracle budget f -cost expansions 2 11.2 2 4 13.5

  47. IBEX (main idea): With an oracle budget f -cost expansions 2 11.2 2 4 13.5 4

  48. IBEX (main idea): With an oracle budget f -cost expansions 2 11.2 2 4 13.5 4 8 13.9 8

  49. IBEX (main idea): With an oracle budget f -cost expansions 2 11.2 2 4 13.5 4 8 13.9 8 16 14.4 16

  50. IBEX (main idea): With an oracle budget f -cost expansions 2 11.2 2 4 13.5 4 8 13.9 8 16 14.4 16 32 16.8 30

  51. IBEX (main idea): With an oracle budget f -cost expansions 2 11.2 2 4 13.5 4 8 13.9 8 16 14.4 16 32 16.8 30 64 17.7 62

  52. IBEX (main idea) For budget = 2, 4, 8, 16, ...: cost_bound = Oracle(budget) search within cost_bound and within budget

  53. IBEX (main idea) For budget = 2, 4, 8, 16, ...: cost_bound = Oracle(budget) search within cost_bound and within budget Oracle(budget) returns largest cost for which budget is su ffi cient. Oracle(budget) = (budgeted) exponential search

  54. IBEX (main idea)

  55. IBEX (main idea)

  56. IBEX (main idea)

  57. IBEX (main idea)

  58. IBEX (main idea)

  59. IBEX (main idea)

  60. IBEX (main idea)

  61. IBEX (main idea)

  62. IBEX (main idea)

  63. IBEX (main idea) N ↓ For budget = 2, 4, 8, 16, 32, 64 : ≤ 4 N  With exponential search :  Refine lower and upper bound on cost  O (log C ∗ ) Proves no solution when lower = upper

  64. Application: Tree Search IBEX + Depth-First Search = Budgeted Tree Search (BTS) Alg. Worst case Θ ( N 2 ) IDA* (1985) EDA* (2014) unbounded BTS O ( N log C ∗ ) C ∗ : cost of the optimal solution N : number of nodes of cost at most C ∗ All algorithms return an optimal solution (under usual assumptions).

  65. Application: Graph Search IBEX + Uniform-Cost Search = Budgeted Graph Search (BGS) Alg. Admissible Heuristic Consistent Inconsistent Ω (2 N ) A* (1968) O ( N ) Θ ( N 2 ) B (1977) O ( N ) O ( N ) † BGS O ( N log C ∗ ) C ∗ : cost of the optimal solution N : number of nodes of cost at most C ∗ All algorithms return an optimal solution (under usual assumptions).

  66. Enhancements: Tree search Recover IDA* when exponential unit-cost domains 1) Infinite iteration at next f -cost 2) If expansions grow exponentially, go to 1) else exponential search

  67. Enhancements: Tree search Recover IDA* when exponential unit-cost domains 1) Infinite iteration at next f -cost 2) If expansions grow exponentially, go to 1) else exponential search Budget window [2 , 8] to be more aggressive

  68. Experiments Algorithm 15-Puzzle (unit) 15-Puzzle (real) Coconut Exp. × 10 6 Exp. × 10 6 Exp. × 10 4 Solved Solved Solved BTS 100 242.5 100 673.1 100 84.7 EDA* 99 5 586.0 100 2 882.3 3 ≥ 554 249 . 0 IDA* CR 100 868.4 100 700.6 3 ≥ 516 937 . 7 IDA* 100 242.5 57 62 044.3 100 5 484.2 100 100.8 100 258.1 100 2.7 N BTS (Budgeted Tree Search) = IBEX + Depth-First Search

  69. Experiments Algorithm 15-Puzzle (unit) 15-Puzzle (real) Coconut Exp. × 10 6 Exp. × 10 6 Exp. × 10 4 Solved Solved Solved BTS 100 242.5 100 673.1 100 84.7 EDA* 99 5 586.0 100 2 882.3 3 ≥ 554 249 . 0 IDA* CR 100 868.4 100 700.6 3 ≥ 516 937 . 7 IDA* 100 242.5 57 62 044.3 100 5 484.2 100 100.8 100 258.1 100 2.7 N BTS (Budgeted Tree Search) = IBEX + Depth-First Search

  70. Experiments Algorithm 15-Puzzle (unit) 15-Puzzle (real) Coconut Exp. × 10 6 Exp. × 10 6 Exp. × 10 4 Solved Solved Solved BTS 100 242.5 100 673.1 100 84.7 EDA* 99 5 586.0 100 2 882.3 3 ≥ 554 249 . 0 IDA* CR 100 868.4 100 700.6 3 ≥ 516 937 . 7 IDA* 100 242.5 57 62 044.3 100 5 484.2 100 100.8 100 258.1 100 2.7 N BTS (Budgeted Tree Search) = IBEX + Depth-First Search

  71. Conclusion IBEX: Budgeting → O ( N log C ) Tree search: drop-in replacement for IDA* DovIBEX: For applications where oracle returns only if a solution is found Interleaves budget iterations Do you have a problem where the current worst case is O ( N 2 ), due to being cautious? Maybe (Dov)IBEX can help! Come and see our poster!

Recommend

A Guide to Budgeted Tree Search Nathan R. Sturtevant University of Alberta Amii Fellow, CIFAR

A Guide to Budgeted Tree Search Nathan R. Sturtevant University of Alberta Amii Fellow, CIFAR

A Guide to Budgeted Tree Search Nathan R. Sturtevant University of Alberta Amii Fellow, CIFAR Chair Malte Helmert Universitt Basel Talk Overview Budgeted Tree Search (BTS) is a new algorithm with better worst-case guarantees than IDA*

1.37k views • 132 slides
Foundations of Artificial Intelligence 12. State-Space Search: Depth-first Search & Iterative

Foundations of Artificial Intelligence 12. State-Space Search: Depth-first Search & Iterative

Foundations of Artificial Intelligence 12. State-Space Search: Depth-first Search & Iterative Deepening Malte Helmert and Thomas Keller University of Basel March 18, 2020 Depth-first Search Iterative Deepening Summary State-Space

1.16k views • 54 slides
Beyond Classical Search Sections 4.1 and 4.2 Ch. 04 p.1/20 Outline Iterative improvement

Beyond Classical Search Sections 4.1 and 4.2 Ch. 04 p.1/20 Outline Iterative improvement

Beyond Classical Search Sections 4.1 and 4.2 Ch. 04 p.1/20 Outline Iterative improvement algorithms Hill climbing Simulated annealing Local beam search Genetic algorithms References: The slides were adapted to the 3 rd edition of Russell

383 views • 20 slides
Search Summary Search Summary Some material from: D Lin, J You, JC Latombe 1 Search Summary #

Search Summary Search Summary Some material from: D Lin, J You, JC Latombe 1 Search Summary #

RN, Chapter 25 Search Summary Search Summary Some material from: D Lin, J You, JC Latombe 1 Search Summary # 1 Problem Solving as Search Blind Search Techniques Breadth-first (uniform cost) Depth-first Iterative

120 views • 8 slides
Stochastic Local Search Methods Dynamic Local Search Iterated Local Search Tabu Search Marco

Stochastic Local Search Methods Dynamic Local Search Iterated Local Search Tabu Search Marco

Outline DM811 HEURISTICS AND LOCAL SEARCH ALGORITHMS FOR COMBINATORIAL OPTIMZATION 1. Stochastic Local Search Methods (Metaheuristics) Randomized Iterative Improvement Lecture 12 Attribute Based Hill Climber Stochastic Local Search Methods

590 views • 9 slides
Uninformed Search Depth First Search Iterative Deepening Volker Sorge Uniform Cost Search

Uninformed Search Depth First Search Iterative Deepening Volker Sorge Uniform Cost Search

Intro to AI: Lecture 2 Volker Sorge States and Actions Background for Search Breadth First Search Uninformed Search Depth First Search Iterative Deepening Volker Sorge Uniform Cost Search Intro to AI: States Lecture 2 Volker Sorge

505 views • 26 slides
Exponential Growth Exponential Growth Introduction Exponential Growth vs. Linear Growth

Exponential Growth Exponential Growth Introduction Exponential Growth vs. Linear Growth

Slide 1 / 175 Slide 2 / 175 Algebra I Algebraic Exponents & Exponential Functions 2015-11-02 www.njctl.org Slide 3 / 175 Slide 4 / 175 Table of Contents Click on the topic to go to that section Exponential Growth Exponential Growth

481 views • 30 slides
Iterative improvement algorithms Prof. Tuomas Sandholm Carnegie Mellon University Computer

Iterative improvement algorithms Prof. Tuomas Sandholm Carnegie Mellon University Computer

Iterative improvement algorithms Prof. Tuomas Sandholm Carnegie Mellon University Computer Science Department Iterative improvement algorithms = iterative refinement = local search Usable when the solution are states, not paths. Start with a

282 views • 10 slides
Depth-First Iterative-Deepening: An Optimal Admissible Tree Search by R. E. Korf Tsan-sheng Hsu

Depth-First Iterative-Deepening: An Optimal Admissible Tree Search by R. E. Korf Tsan-sheng Hsu

Depth-First Iterative-Deepening: An Optimal Admissible Tree Search by R. E. Korf Tsan-sheng Hsu tshsu@iis.sinica.edu.tw http://www.iis.sinica.edu.tw/~tshsu 1 Abstract The complexities of various search algorithms are considered in terms of

1.04k views • 41 slides
Depth-Bound Heuristics and Iterative-Deepening Search Algorithms in Classical Planning

Depth-Bound Heuristics and Iterative-Deepening Search Algorithms in Classical Planning

Depth-Bound Heuristics and Iterative-Deepening Search Algorithms in Classical Planning Bachelors Thesis Presentation Florian Spiess, 13 June 2017 Departement of Mathematics and Computer Science Artificial Intelligence Classical Planning

339 views • 33 slides
V OCABULARY : Exponential growth This is a situation where something increases by a constant

V OCABULARY : Exponential growth This is a situation where something increases by a constant

D AY 113 W RITE THE EQUATION OF AN EXPONENTIAL GROWTH OF A FUNCTION (D AY 1) I NTRODUCTION We exhibit many situations involving exponential growth of functions. For instance, the increase in the amounts of investment is a case of exponential

404 views • 11 slides
Local Search [These slides were created by Dan Klein and Pieter Abbeel for CS188 Intro to AI at UC

Local Search [These slides were created by Dan Klein and Pieter Abbeel for CS188 Intro to AI at UC

Local Search [These slides were created by Dan Klein and Pieter Abbeel for CS188 Intro to AI at UC Berkeley. All CS188 materials are available at http://ai.berkeley.edu.] Iterative Improvement Iterative Algorithms for CSPs Local search

421 views • 13 slides
Search-based Testing of Procedural Programs: Iterative Single-Target or Multi-Target Approach?

Search-based Testing of Procedural Programs: Iterative Single-Target or Multi-Target Approach?

Search-based Testing of Procedural Programs: Iterative Single-Target or Multi-Target Approach? Simone Scalabrino Giovanni Grano Dario Di Nucci Rocco Oliveto Andrea De Lucia The overall cost of testing has been estimated at being at

568 views • 39 slides
Uninformed Search Alice Gao Lecture 3 Based on work by K. Leyton-Brown, K. Larson, and P. van

Uninformed Search Alice Gao Lecture 3 Based on work by K. Leyton-Brown, K. Larson, and P. van

1/34 Uninformed Search Alice Gao Lecture 3 Based on work by K. Leyton-Brown, K. Larson, and P. van Beek 2/34 Outline Learning Goals Uninformed Search Breadth-First Search Depth-First Search Iterative-Deepening Search Comparing the

732 views • 34 slides
1 2 Total Budgeted in Total Planned in Total Planned in 2019 for Total Budgeted 2020 for

1 2 Total Budgeted in Total Planned in Total Planned in 2019 for Total Budgeted 2020 for

1 2 Total Budgeted in Total Planned in Total Planned in 2019 for Total Budgeted 2020 for 2021 for Total Budgeted for Carryover/Multi- in 2019 for New Total Budgeted Carryover & Carryover & 2018 year Projects Projects for

406 views • 7 slides
X Example? In such cases, we can use local search algorithms Keep a single

X Example? In such cases, we can use local search algorithms Keep a single

Todays 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

236 views • 5 slides
Iterative Search for Weakly Supervised Semantic Parsing Pradeep Matt Shikhar Luke Ed Hovy

Iterative Search for Weakly Supervised Semantic Parsing Pradeep Matt Shikhar Luke Ed Hovy

Iterative Search for Weakly Supervised Semantic Parsing Pradeep Matt Shikhar Luke Ed Hovy Dasigi Gardner Murty Zettlemoyer This talk in one slide Training semantic parsing with denotation-only supervision is challenging because of

229 views • 21 slides
Metaheuristics 2.3 Local Search 2.4 Simulated annealing Adrian Horga 1 2.3 Local Search 2

Metaheuristics 2.3 Local Search 2.4 Simulated annealing Adrian Horga 1 2.3 Local Search 2

Metaheuristics 2.3 Local Search 2.4 Simulated annealing Adrian Horga 1 2.3 Local Search 2 Local Search Other names: Hill climbing Descent Iterative improvement General S-Metaheuristics Old and simple method at each

620 views • 25 slides
recursion 2: power, binary search Oct. 4/5, 2017 1 Recall: Converting to binary (iterative)

recursion 2: power, binary search Oct. 4/5, 2017 1 Recall: Converting to binary (iterative)

COMP 250 Lecture 12 recursion 2: power, binary search Oct. 4/5, 2017 1 Recall: Converting to binary (iterative) k = 0 while n > 0 { b[ k ] = n % 2 n = n / 2 k++ } Recall that in binary needs approximately bits.

663 views • 29 slides
Introduction to exponential functions An exponential function is a function of the form f ( x ) = b

Introduction to exponential functions An exponential function is a function of the form f ( x ) = b

Introduction to exponential functions An exponential function is a function of the form f ( x ) = b x where b is a Elementary Functions fixed positive number. The constant b is called the base of the exponent. Part 3, Exponential Functions &

776 views • 6 slides
Authority Financial Snapshot May 2016 Profit/Loss $746,108 $99,449 ($386,772) Budgeted

Authority Financial Snapshot May 2016 Profit/Loss $746,108 $99,449 ($386,772) Budgeted

June 28, 2016 Board of Governors Meeting Projects Finance Authority Financial Snapshot May 2016 Profit/Loss $746,108 $99,449 ($386,772) Budgeted $1,271,277 ($2,199,145) $474,014 Actual $1,931,285 $27,685,988 $99,386 Budgeted

795 views • 41 slides
Applications of exponential functions Applications of exponential functions abound throughout the

Applications of exponential functions Applications of exponential functions abound throughout the

Applications of exponential functions Applications of exponential functions abound throughout the sciences. Exponential functions are the primary functions that scientists work with. Here are some examples. Elementary Functions Exponential

241 views • 7 slides
Outline Problem solving agents and search Solving Problems by Examples Properties

Outline Problem solving agents and search Solving Problems by Examples Properties

Outline Problem solving agents and search Solving Problems by Examples Properties of search algorithms Searching Uninformed search CS486/686 Breadth first Depth first University of Waterloo Iterative Deepening

413 views • 6 slides
Improved Approximation Algorithms for Budgeted Allocations Lecture outlines 1. Presenting the

Improved Approximation Algorithms for Budgeted Allocations Lecture outlines 1. Presenting the

Improved Approximation Algorithms for Budgeted Allocations Lecture outlines 1. Presenting the budgeted allocation problem 2. Defining some basic terms we are going to use 3. The 3/2-approximation algorithm 4. Analyzing its correctness and

738 views • 59 slides

More recommend