140 likes | 255 Views
Trivial Parallelization of an Existing EA. Asher Freese CS401. Motivation. Many difficult problems NP-Complete 0/1 Knapsack Traveling Salesman Initial Solution: Evolutionary Algorithm Problem: EAs rarely get optimal solution
E N D
Trivial Parallelization of an Existing EA Asher Freese CS401
Motivation • Many difficult problems • NP-Complete • 0/1 Knapsack • Traveling Salesman • Initial Solution: Evolutionary Algorithm • Problem: • EAs rarely get optimal solution • Some problems are even bigger than an EA can handle in reasonable time
Solution • Parallel Computing (Island Model) • Parallelization increases speed of programs • EAs are inherently parallelizable • Easy to split populations • Most suitable uses • Requires many generations to converge • Expensive fitness function
Previous Research • Very Little • Proof of concept • Speed up computation of unrelated problem • Method of achieving population diversity
Approach • Parallel, Island Model framework • Parallel framework takes EA class • Instantiation by main will not change • Loop • Run for some generations • Trade individuals among processes • EA class inherits parent for parallel framework
Two types of Parallelization • Purely Parellel • Each EA is completely independent • Island Model trading • EA’s mostly independent • Trade individuals occasionally
Testing • N-queens problem • 0/1 Knapsack (later) • Ackley Function (later)
Preliminary Conclusion • Low epoch length, bad • Low number traded individuals, good • High number of processors, good
Future Work • More tests • 0/1 Knapsack • Ackley Function • Different trading algorithms • Processor niching