online motion planning ma-inf 1314 introduction filedfs s s optimal online motion...
TRANSCRIPT
![Page 1: Online Motion Planning MA-INF 1314 Introduction fileDFS s s Optimal Online Motion PlanningIntroduction 18.4.2017 c Elmar LangetepeSS ’178. Online motion planning: Modell Searching](https://reader031.vdocuments.site/reader031/viewer/2022041301/5e1056f9bd69fa0726311746/html5/thumbnails/1.jpg)
Online Motion Planning MA-INF 1314Introduction
Elmar Langetepe
University of Bonn
Online Motion Planning Introduction 18.4.2017 c©Elmar Langetepe SS ’17 1
![Page 2: Online Motion Planning MA-INF 1314 Introduction fileDFS s s Optimal Online Motion PlanningIntroduction 18.4.2017 c Elmar LangetepeSS ’178. Online motion planning: Modell Searching](https://reader031.vdocuments.site/reader031/viewer/2022041301/5e1056f9bd69fa0726311746/html5/thumbnails/2.jpg)
Motion planning for mobile agents
• Lecture: Tuesday/Thursday 10:15 to 11:45
• Exercise groups: Starting next week 27th
Tuesday/Thursday: 12-14
• Sign in now
• Manuscript on the webpage
• Slides on the webpage
• Exercises
• Today: Short intro, different topics, some examples
Online Motion Planning Introduction 18.4.2017 c©Elmar Langetepe SS ’17 2
![Page 3: Online Motion Planning MA-INF 1314 Introduction fileDFS s s Optimal Online Motion PlanningIntroduction 18.4.2017 c Elmar LangetepeSS ’178. Online motion planning: Modell Searching](https://reader031.vdocuments.site/reader031/viewer/2022041301/5e1056f9bd69fa0726311746/html5/thumbnails/3.jpg)
Motion planning categories
• Elektronic devices
• Mechanical devices
• Control/Process engineering
• Artificial Intelligence
• Softwareengineering
• ...
• Plans: Algorithmic
• Input: Geometry of the Environment
Online Motion Planning Introduction 18.4.2017 c©Elmar Langetepe SS ’17 3
![Page 4: Online Motion Planning MA-INF 1314 Introduction fileDFS s s Optimal Online Motion PlanningIntroduction 18.4.2017 c Elmar LangetepeSS ’178. Online motion planning: Modell Searching](https://reader031.vdocuments.site/reader031/viewer/2022041301/5e1056f9bd69fa0726311746/html5/thumbnails/4.jpg)
Geometric Algorithms
• Geometry of Agent and Environment
• Solve motion planning problem
• Main difference: Online/Offline
• Incomplete Information
• Correctnes? Efficiency?
• Lower/Upper Bound
• Structural Properties
• Groundtasks
• Algorithms paradigms
• Formal proofs
Online Motion Planning Introduction 18.4.2017 c©Elmar Langetepe SS ’17 4
![Page 5: Online Motion Planning MA-INF 1314 Introduction fileDFS s s Optimal Online Motion PlanningIntroduction 18.4.2017 c Elmar LangetepeSS ’178. Online motion planning: Modell Searching](https://reader031.vdocuments.site/reader031/viewer/2022041301/5e1056f9bd69fa0726311746/html5/thumbnails/5.jpg)
Example I: Simple polygon, exploration
• Simple room, no obstacles
• Agent starts at s, sees everythings, returns to the start
• 360 degree vision
• Modell: Point-shaped agent, simple polygons, visibility polygon,
• Optimal route: Shortest Watchman Route
• Offline: Polygon is fully known
• Online: Union of visibility polygons
• Offline: Optimal algorithm
• Applets: Offline algorithm versus Online algorithm
Online Motion Planning Introduction 18.4.2017 c©Elmar Langetepe SS ’17 5
![Page 6: Online Motion Planning MA-INF 1314 Introduction fileDFS s s Optimal Online Motion PlanningIntroduction 18.4.2017 c Elmar LangetepeSS ’178. Online motion planning: Modell Searching](https://reader031.vdocuments.site/reader031/viewer/2022041301/5e1056f9bd69fa0726311746/html5/thumbnails/6.jpg)
Example I: Online/Offline
• Axisparallel polygon
• Offline Algorithm: O(n) Applet!
• Online, not optimal!
• Greedy Online Strategy!
• Proof: L1-optimal!
•√
2 Approximation!
• Competitive Ratio
Online Motion Planning Introduction 18.4.2017 c©Elmar Langetepe SS ’17 6
![Page 7: Online Motion Planning MA-INF 1314 Introduction fileDFS s s Optimal Online Motion PlanningIntroduction 18.4.2017 c Elmar LangetepeSS ’178. Online motion planning: Modell Searching](https://reader031.vdocuments.site/reader031/viewer/2022041301/5e1056f9bd69fa0726311746/html5/thumbnails/7.jpg)
Example II: Exploration of grid-environments
• Grid world (cells), tool of cell-size, one-step from cell to cell
• Process any cell in the connected-component
• Touchsensor only detects neighboring cells
• Build a map, visit all cells, return to start
• Grid-explore applet
s ??
??s
Online Motion Planning Introduction 18.4.2017 c©Elmar Langetepe SS ’17 7
![Page 8: Online Motion Planning MA-INF 1314 Introduction fileDFS s s Optimal Online Motion PlanningIntroduction 18.4.2017 c Elmar LangetepeSS ’178. Online motion planning: Modell Searching](https://reader031.vdocuments.site/reader031/viewer/2022041301/5e1056f9bd69fa0726311746/html5/thumbnails/8.jpg)
Example: Exploration of grid-environments
• Online DFS for cells ⇒: Strategy
• Visit all connected cells: Correctness
• Number of steps: 2× (C − 1)
• Optimum: At least C
• Ratio: smaller than 2 times the optimum, Performance
• No better strategy! Lower Bound!!
• Provoke detour! Proof of this later!!!
DFS
s s
Optimal
Online Motion Planning Introduction 18.4.2017 c©Elmar Langetepe SS ’17 8
![Page 9: Online Motion Planning MA-INF 1314 Introduction fileDFS s s Optimal Online Motion PlanningIntroduction 18.4.2017 c Elmar LangetepeSS ’178. Online motion planning: Modell Searching](https://reader031.vdocuments.site/reader031/viewer/2022041301/5e1056f9bd69fa0726311746/html5/thumbnails/9.jpg)
Online motion planning: Modell
• Searching for a goal
• Exploration of an environment
• Process task on (subset of) environment
• Escape from a labyrinth
• Continuous/discrete vision
• Touch sensor/compass
• Build a map
• History: Simple ⇒ complicated
Online Motion Planning Introduction 18.4.2017 c©Elmar Langetepe SS ’17 9
![Page 10: Online Motion Planning MA-INF 1314 Introduction fileDFS s s Optimal Online Motion PlanningIntroduction 18.4.2017 c Elmar LangetepeSS ’178. Online motion planning: Modell Searching](https://reader031.vdocuments.site/reader031/viewer/2022041301/5e1056f9bd69fa0726311746/html5/thumbnails/10.jpg)
Competitive analysis, competitive ratio
Definition 1: Let Π be a problem class and S be a strategy, that
solves any instance P ∈ Π.
Let KS(P ) be the cost of S for solving P .
Let KOPT(P ) be the cost of the optimal solution for P .
The strategy S is denoted to be c–competitive, if there are fixed
constants c, α > 0, so that for all P ∈ Π
KS(P ) ≤ c ·KOPT(P ) + α
holds.
Online Motion Planning Introduction 18.4.2017 c©Elmar Langetepe SS ’17 10
![Page 11: Online Motion Planning MA-INF 1314 Introduction fileDFS s s Optimal Online Motion PlanningIntroduction 18.4.2017 c Elmar LangetepeSS ’178. Online motion planning: Modell Searching](https://reader031.vdocuments.site/reader031/viewer/2022041301/5e1056f9bd69fa0726311746/html5/thumbnails/11.jpg)
Chap. 1: Labyrinths, Grids, Graphs
• Different 2D environments
• Tasks: Searching for a goal, escape, exploration
• Def: Labyrinth L intuitiv: Divide the plane by walls into corridors
• Def: Grid-Labyrinth grid/cell enviroment with walls on the edges
• History 1950: Shannon 5× 5 Labyrinth with an elektr. Mouse
T
S
Online Motion Planning Introduction 18.4.2017 c©Elmar Langetepe SS ’17 11
![Page 12: Online Motion Planning MA-INF 1314 Introduction fileDFS s s Optimal Online Motion PlanningIntroduction 18.4.2017 c Elmar LangetepeSS ’178. Online motion planning: Modell Searching](https://reader031.vdocuments.site/reader031/viewer/2022041301/5e1056f9bd69fa0726311746/html5/thumbnails/12.jpg)
Chap. 1.1 Shannons Mouse Alg.
Historie 1950: Shannon 5× 5 Labyrinth with an elektr. Mouse
• For any cell, there is a marker that stores the direction (N,E,S,W)
in which the mouse left the cell at the last visit.
• Initialize any cell with marker N (north)
• While T has not been visited:
Search for the first free cell in clockwise order starting from the
current marker direction.
Change the marker correspondingly and enter the corresponding
cell.
Online Motion Planning Introduction 18.4.2017 c©Elmar Langetepe SS ’17 12
![Page 13: Online Motion Planning MA-INF 1314 Introduction fileDFS s s Optimal Online Motion PlanningIntroduction 18.4.2017 c Elmar LangetepeSS ’178. Online motion planning: Modell Searching](https://reader031.vdocuments.site/reader031/viewer/2022041301/5e1056f9bd69fa0726311746/html5/thumbnails/13.jpg)
Shannons Maus Alg.
T
S
NS
NO NO N
Online Motion Planning Introduction 18.4.2017 c©Elmar Langetepe SS ’17 13
![Page 14: Online Motion Planning MA-INF 1314 Introduction fileDFS s s Optimal Online Motion PlanningIntroduction 18.4.2017 c Elmar LangetepeSS ’178. Online motion planning: Modell Searching](https://reader031.vdocuments.site/reader031/viewer/2022041301/5e1056f9bd69fa0726311746/html5/thumbnails/14.jpg)
Shannons Mouse: Correctness!
Theorem 1.1: Shannons Mouse Algorithm will always find the
target in any grid labyrinth from any starting point, if a path from S
to T exists.
Proof:
• Consider the strategy without the target
• Show: All reachable cells will be visited infinitely often
• Formal proof: Blackboard!
Online Motion Planning Introduction 18.4.2017 c©Elmar Langetepe SS ’17 14
![Page 15: Online Motion Planning MA-INF 1314 Introduction fileDFS s s Optimal Online Motion PlanningIntroduction 18.4.2017 c Elmar LangetepeSS ’178. Online motion planning: Modell Searching](https://reader031.vdocuments.site/reader031/viewer/2022041301/5e1056f9bd69fa0726311746/html5/thumbnails/15.jpg)
Efficient algorithm for graph-exploration
• Graph-exploration, visit all edges (and vertices)
• Vertex: Outgoing edges become visible! Build the full graph,
return to the start!
• Visited edges can be located
• Strategy: Online-DFS for the edges, visits any edges twice
Theorem: Exploring an unknown graph requires roughly twice as
many edge visits than the optimal exploration route for the known
graph. DFS requires no more that twice as many edges.
Formal proof! Second part is already clear! Lower bound by
worst-case adversary strategy
Online Motion Planning Introduction 18.4.2017 c©Elmar Langetepe SS ’17 15
![Page 16: Online Motion Planning MA-INF 1314 Introduction fileDFS s s Optimal Online Motion PlanningIntroduction 18.4.2017 c Elmar LangetepeSS ’178. Online motion planning: Modell Searching](https://reader031.vdocuments.site/reader031/viewer/2022041301/5e1056f9bd69fa0726311746/html5/thumbnails/16.jpg)
Graphexploration, Edge visits, Adversary
Adversary: 2− δ worse than the optimum
corridor, agent from s has explored l vertices
s
l
Now bifurcation at s′
l1
l
s′
s
Online Motion Planning Introduction 18.4.2017 c©Elmar Langetepe SS ’17 16
![Page 17: Online Motion Planning MA-INF 1314 Introduction fileDFS s s Optimal Online Motion PlanningIntroduction 18.4.2017 c Elmar LangetepeSS ’178. Online motion planning: Modell Searching](https://reader031.vdocuments.site/reader031/viewer/2022041301/5e1056f9bd69fa0726311746/html5/thumbnails/17.jpg)
Graphexploration, Edge visits, Adversary
• Case 1: Agent returns to s
l2
ss′
l3
l1
• Case 2: Agent reaches the end of the start corridor
l1
l
l2
ss′
• Close the corridors, present the scene
Online Motion Planning Introduction 18.4.2017 c©Elmar Langetepe SS ’17 17
![Page 18: Online Motion Planning MA-INF 1314 Introduction fileDFS s s Optimal Online Motion PlanningIntroduction 18.4.2017 c Elmar LangetepeSS ’178. Online motion planning: Modell Searching](https://reader031.vdocuments.site/reader031/viewer/2022041301/5e1056f9bd69fa0726311746/html5/thumbnails/18.jpg)
Adversary analysis: Case 1.
• Strategy SROB against SOPT
• Case 1: Agent returns to s
• Up to now |SROB| ≥ 2l1 + 2l2 + 2l3 + 2(l − l1) = 2(l + l2 + l3)
• From now on: 2(l + l2 + l3) + 6
• Optimal: |SOPT| = 2(l + l2 + l3) + 6
• |SROB| ≥ (2− δ)|SOPT| for l ≥⌈3δ
⌉
l2
ss′
l3
l1
Online Motion Planning Introduction 18.4.2017 c©Elmar Langetepe SS ’17 18
![Page 19: Online Motion Planning MA-INF 1314 Introduction fileDFS s s Optimal Online Motion PlanningIntroduction 18.4.2017 c Elmar LangetepeSS ’178. Online motion planning: Modell Searching](https://reader031.vdocuments.site/reader031/viewer/2022041301/5e1056f9bd69fa0726311746/html5/thumbnails/19.jpg)
Adversary analysis: Case 2.
• Case 2: Agent reaches the end of start corridor
• Up to now: |SROB| ≥ 2l1 + l − l1 + 2l2 + l + 1
• From now on: l + 1 + 2(l2 + 1) + l − l1• Sum: |SROB| = 4(l + l2) + 4
• Optimal: |SOPT| = 2(l + 1 + l2 + 1) = 2(l + l2) + 4
• |SROB| ≥ (2− δ)|SOPT| fur l ≥⌈3δ
⌉
l1
l
l2
ss′
Online Motion Planning Introduction 18.4.2017 c©Elmar Langetepe SS ’17 19
![Page 20: Online Motion Planning MA-INF 1314 Introduction fileDFS s s Optimal Online Motion PlanningIntroduction 18.4.2017 c Elmar LangetepeSS ’178. Online motion planning: Modell Searching](https://reader031.vdocuments.site/reader031/viewer/2022041301/5e1056f9bd69fa0726311746/html5/thumbnails/20.jpg)
Remarks
Corollary : DFS for Online-Exploration of graphs is 2-competitiv,
there is no C-competitive strategy with C < 2.
• DFS is optimal
• Additive constant: Start situation
• Example: Goal very close
• Lower bounds only with arbitrary large examples
Online Motion Planning Introduction 18.4.2017 c©Elmar Langetepe SS ’17 20
![Page 21: Online Motion Planning MA-INF 1314 Introduction fileDFS s s Optimal Online Motion PlanningIntroduction 18.4.2017 c Elmar LangetepeSS ’178. Online motion planning: Modell Searching](https://reader031.vdocuments.site/reader031/viewer/2022041301/5e1056f9bd69fa0726311746/html5/thumbnails/21.jpg)
Remarks!
• Return to the target
• No return: Exercise
• Examples: Opt. Tour visits any edge twice
• Visit only the vertices, DFS for vertices? Exercise!
Online Motion Planning Introduction 18.4.2017 c©Elmar Langetepe SS ’17 21