400 likes | 599 Views
Finding Optimal Solutions to Cooperative Pathfinding Problems. Trevor Standley Computer Science Department University of California, Los Angeles http://cs.ucla.edu/~tstand/. Introduction. Pathfinding Problems A single agent must find a path from a start state to a goal state
E N D
Finding Optimal Solutions to Cooperative Pathfinding Problems Trevor Standley Computer Science Department University of California, Los Angeles http://cs.ucla.edu/~tstand/
Introduction • Pathfinding Problems • A single agent must find a path from a start state to a goal state • Cooperative Pathfinding Problems • Multiple agents interact • Want to minimize the total cost
My Formulation • Gridworld pathfinding
Related Work • Centralized Approaches • Strengths: Typically complete, can be optimal • Weaknesses: Takes forever! • Decoupled Approaches • Strengths: Fast • Weaknesses: Incomplete and suboptimal
Related Work • Centralized Approaches • Strengths: Typically complete, can be optimal • Weaknesses: Takes forever! • Decoupled Approaches • Strengths: Fast • Weaknesses: Incomplete and suboptimal
Related Work • Centralized Approaches • Strengths: Typically complete, can be optimal • Weaknesses: Takes forever! • Decoupled Approaches • Strengths: Fast • Weaknesses: Incomplete and suboptimal
Related Work • Centralized Approaches • Strengths: Typically complete, can be optimal • Weaknesses: Takes forever! • Decoupled Approaches • Strengths: Fast • Weaknesses: Incomplete and suboptimal
Related Work • Centralized Approaches • Strengths: Typically complete, can be optimal • Weaknesses: Takes forever! • Decoupled Approaches • Strengths: Fast • Weaknesses: Incomplete and suboptimal
Related Work • Centralized Approaches • Strengths: Typically complete, can be optimal • Weaknesses: Takes forever! • Decoupled Approaches • Strengths: Fast • Weaknesses: Incomplete and suboptimal
Related Work • Centralized Approaches • Strengths: Typically complete, can be optimal • Weaknesses: Takes forever! • Decoupled Approaches • Strengths: Fast • Weaknesses: Incomplete and suboptimal
Related Work • Centralized Approaches • Strengths: Typically complete, can be optimal • Weaknesses: Takes forever! • Decoupled Approaches • Strengths: Fast • Weaknesses: Incomplete and suboptimal
Related Work • Centralized Approaches • Strengths: Typically complete, can be optimal • Weaknesses: Takes forever! • Decoupled Approaches • Strengths: Fast • Weaknesses: Incomplete and suboptimal
Related Work • Centralized Approaches • Strengths: Typically complete, can be optimal • Weaknesses: Takes forever! • Decoupled Approaches • Strengths: Fast • Weaknesses: Incomplete and suboptimal
Related Work • Centralized Approaches • Strengths: Typically complete, can be optimal • Weaknesses: Takes forever! • Decoupled Approaches • Strengths: Fast • Weaknesses: Incomplete and suboptimal
Related Work • Centralized Approaches • Strengths: Typically complete, can be optimal • Weaknesses: Takes forever! • Decoupled Approaches • Strengths: Fast • Weaknesses: Incomplete and suboptimal
Related Work • Centralized Approaches • Strengths: Typically complete, can be optimal • Weaknesses: Takes forever! • Decoupled Approaches • Strengths: Fast • Weaknesses: Incomplete and suboptimal
Related Work • Centralized Approaches • Strengths: Typically complete, can be optimal • Weaknesses: Takes forever! • Decoupled Approaches • Strengths: Fast • Weaknesses: Incomplete and suboptimal
Related Work • Centralized Approaches • Strengths: Typically complete, can be optimal • Weaknesses: Takes forever! • Decoupled Approaches • Strengths: Fast • Weaknesses: Incomplete and suboptimal
The Standard Algorithm • The standard algorithm is A* • Centralized algorithm • There is a standard heuristic • State representation – A position for each agent • State space – Exponential in the number of agents • An operator – Complete assignment of moves to agents • One of {N; NE; E; SE; S; SW; W; NW; and wait} for each agent • Exponential in the number of agents • Obviously this algorithm is not taken seriously
My algorithm • Optimal • Complete • Two main contributions • Operator decomposition • Independence detection
Operator Decomposition • Intuition • Also a centralized algorithm • Still use A* • Change how operators are defined: only one agent moves at a time • Simple idea, tricky to get details right
Operator Decomposition • Each operator assigns a move to a single agent • Assignments are made in a fixed order • Move assignments stored as part of the state representation
Operator Decomposition • Example
Consequences of Operator Decomposition • Branching factor becomes polynomial • However, state space still exponential
Simple Independence Detection • Create a group for each agent • Plan paths for each group independently • Check for conflicts in new paths • Combine groups with conflicting paths • Repeat 2-4 until no conflicts
Simple Independence Detection Problem • Are these agents independent?
Simple Independence Detection Problem • Are these agents independent?
Better Independence Detection • When a conflict is detected between two groups, try to find an alternate path for one of the groups • If that fails try to find an alternate path for the other group • Only combine groups if no alternate path could be found
Independence Detection • Which alternate paths are the best? • Only search for optimal paths • Paths can be found using operator decomposition • Find paths that will lead to fewest number of future conflicts • Operator decomposition can be modified to find optimal paths with few future conflicts
My Algorithm • Uses decoupled planning where possible • Only uses centralized planning for non-independent subproblems • Calls operator decomposition as a subroutine to do the centralized planning
Results • 10000 randomly generated problems with 2-60 agents
Conclusions • Researchers have developed centralized and decoupled approaches for solving cooperative pathfinding problems • Operator decomposition is an improved centralized approach • Independence detection is a hybrid approach • Only uses centralized planning when necessary
Acknowledgments • My advisor, Rich Korf. • Dawn Chen for editing, advice, and artwork