dan_presentation_notes.txt

From Monday: 15 min talk with 5-10min reserved for Q&A.

1. Describe the problem
2. Describe what has been attempted and not worked.
3. Describe our approach to the problem and why it may work.


I: Problem
	Machine Learning to optimize parameters in genetic algorithms
	Parallelizing the overall process to speed it up
	Interfacing with LAMMPS to determine how good the results are

II: Attempted
	Different Parallel Genetic Algorithms
		Single-population fine-grained
			Suited for massively parallel PCs
			One spatially-structured population
			Selection/mating restricted to small neighborhood
		Global single-population master-slave
		Multiple-population coarse grained
			More sophisticated: consist of subpopulations that exchange individuals (migration)
			Migration is controlled by several parameters

III: Approach
	Crossover of GA traits simulated with binary method (cant98.pdf, pg. 3 of 30)
		90% crossover probability, 10% mutation probability (Project description.pdf)
	Global single-poplulation master-slave GA (cant98.pdf, section 5)
		Master stores the population, executes GA ops, and distributes to slaves
		Slaves only evaluate the fitness of what they've been given (single or subset)
		Star-like approach, one per processor
		Evaluation of fitness is distributed
		Figure 1 from cant98.pdf
	Computation and communication ratio for distribution?
	Master performs GA operations on the dataset to spawn off individuals
	The slave processes fitness of their individuals and feed the values into master
	Synchronous: wait for all results and then move on (since we need to evaluate all individuals)
	Straightforward implementation and doesn't modify how the GA performs its work

Powerpoint slides and notes
	Introduction to PGAs
	Overview of the three PGAs we considered initially
	Overview of LAMMPS
	Diagram of how this all interacts?
	Decision of PGA and why it would be the best
	Dr. Liu: each generation takes about one hour
		Design goals: reduce by a factor of 10?