1. Martin Pelikan and James D. Laury Jr. study the effects of parallelizing the model building process in hBOA on its scalability.
2. They find that parallelizing model building provides nearly linear speedups but can slightly increase the number of function evaluations needed.
3. However, the negative effects of parallelizing model building are negligible compared to the performance gains, so parallelizing model building is beneficial when it is the bottleneck.
Scaling API-first – The story of a global engineering organization
Effect of Parallelizing Model Building in hBOA on Scalability
1. Order or Not: Does Parallelization of Model
Building in hBOA Affect Its Scalability?
Martin Pelikan and James D. Laury, Jr.
Missouri Estimation of Distribution Algorithms Laboratory (MEDAL)
University of Missouri, St. Louis, MO
http://medal.cs.umsl.edu/
pelikan@cs.umsl.edu
Martin Pelikan, James D. Laury, Jr. Parallelization of Model Building in hBOA and Scalability
2. Motivation
Background
hBOA can solve nearly decomposable and hierarchical
problems scalably, often in O(n2 ) evaluations or faster.
O(n2 ) is sometimes not enough...
We may need efficiency enhancemente techniques
Parallelization, evaluation relaxation, hybridization, ...
Parallelization of model building
Among the most powerful efficiency enhancements.
But modifies the model building procedure.
Speedups obtained were analyzed.
But effects of modifying the model building procedure on
hBOA scalability were not analyzed thoroughly.
Purpose
Study the effects of model-building parallelization on hBOA
performance.
Martin Pelikan, James D. Laury, Jr. Parallelization of Model Building in hBOA and Scalability
3. Outline
1. Hierarchical BOA (hBOA) basics.
2. Parallelization of model building in hBOA.
3. Experiments.
4. Summary and conclusions.
Martin Pelikan, James D. Laury, Jr. Parallelization of Model Building in hBOA and Scalability
4. Hierarchical Bayesian Optimization Algorithm (hBOA)
Hierarchical BOA (Pelikan & Goldberg; 2000, 2001)
Build and sample a Bayes net instead of crossover/mutation.
Use local structures and niching.
Current Selected Bayesian New
population population network population
Martin Pelikan, James D. Laury, Jr. Parallelization of Model Building in hBOA and Scalability
5. Bayesian Network
Bayesian network has two parts
Structure
Structure determines edges in the network.
X0
X5 X1
X4 X2
X3
Parameters.
Parameters specify conditional probabilities of each variable
given its parents (variables that this variable depends on).
Example: p(X2 |X0 , X1 ).
Martin Pelikan, James D. Laury, Jr. Parallelization of Model Building in hBOA and Scalability
6. Efficiency Enhancement of hBOA
Potential bottlenecks
1. Model building.
2. Evaluation of candidate solutions.
Efficiency enhancement
1. Parallelization.
2. Hybridization.
3. Time continuation.
4. Fitness evaluation relaxation.
5. Prior knowledge utilization.
6. Incremental and sporadic model building.
7. Learning from experience.
Martin Pelikan, James D. Laury, Jr. Parallelization of Model Building in hBOA and Scalability
7. Parallel hBOA
What to parallelize?
1. Model building.
2. Fitness evaluation.
Parallelizing evaluation
Same situation as for other evolutionary algorithms.
Use master-slave architecture to distribute evaluations.
This work
Focus on parallelization of model building.
Martin Pelikan, James D. Laury, Jr. Parallelization of Model Building in hBOA and Scalability
8. Model Building: Basics
Components of model building
1. Learn structure (dependencies).
2. Learn parameters (conditional probabilities).
Learning structure
Start with an empty network.
Add one edge at a time (without creating cycles).
Choose edges that improve the network the most.
Learning parameters
Compute frequencies from selected solutions.
Estimate probabilities from computed frequencies.
Martin Pelikan, James D. Laury, Jr. Parallelization of Model Building in hBOA and Scalability
9. Complexity Analysis of Model Building
Notation
Upper bound on the number of parents: k
Population size: N
Number of bits: n
Analysis
Learn structure: O(kn2 N ).
Learn parameters: O(knN ).
Bottom line
Learning structure = most expensive part.
Should parallelize the greedy algorithm for learning structure.
Martin Pelikan, James D. Laury, Jr. Parallelization of Model Building in hBOA and Scalability
10. Parallelizing Structure Learning
Master-slave approach (Ocenasek & Schwarz, 2000)
1. Master sends each node to a separate slave processor.
2. Master distributes the population to all slave processors.
3. Each slave processor determines parents for its node.
4. The master collects the results and learns parameters.
Variations
Each processor deals with multiple nodes.
Parameters are also learned locally in slave processors.
Martin Pelikan, James D. Laury, Jr. Parallelization of Model Building in hBOA and Scalability
11. Ordering Nodes to Avoid Cycles
Motivation
Master-slave approach looks great.
But Bayesian network cannot contain cycles!
Communication required after each new edge on any slave.
This leads to a LOT of communication.
Ordering nodes (Ocenasek & Schwarz, 2000)
Order the nodes randomly before learning the structure.
Only allow edges consistent with the ordering.
No need to check for cycles.
Each processor can deal with edges into a single node.
Communication needed only before/after the learning.
Martin Pelikan, James D. Laury, Jr. Parallelization of Model Building in hBOA and Scalability
12. Ordering Nodes: Example
Consider ordering (X1 , X3 , X2 , X4 , X0 , X5 )
Consistent network Inconsistent network
Martin Pelikan, James D. Laury, Jr. Parallelization of Model Building in hBOA and Scalability
13. Example Speedups with Parallel hBOA
Results from Ocenasek et al. (2004)
Nearly linear speedups even for relatively small problems.
No need for parallel computers with shared memory.
Martin Pelikan, James D. Laury, Jr. Parallelization of Model Building in hBOA and Scalability
14. Effects of Node Ordering
Positive effects
Straightforward parallelization with little communication.
Slightly faster model learning even on a sequential computer.
Negative effects
Model structures are restricted.
Models may become less accurate.
Population size and number of generations might increase.
Martin Pelikan, James D. Laury, Jr. Parallelization of Model Building in hBOA and Scalability
15. Description of Experiments
Test functions
Concatenated trap of order 3 and 5.
Hierarchical trap.
Ising spin glass (2D).
Scalability analysis
Number of function evaluations with respect to problem size.
Parameters
Population size from bisection to ensure 30/30 successful runs.
Number of generations upper bounded by n.
Window size in RTR: w = min{n, N/20}.
Martin Pelikan, James D. Laury, Jr. Parallelization of Model Building in hBOA and Scalability
16. Results on dec-3
hBOA
hBOA with ordered nodes
Number of Evaluations
100000
10000
30 60 90 120 150
Problem Size
hBOA with node ordering needs slightly more evaluations.
But differences are relatively small.
Martin Pelikan, James D. Laury, Jr. Parallelization of Model Building in hBOA and Scalability
17. Results on trap-5
1e+06
hBOA
hBOA with ordered nodes
Number of Evaluations
100000
10000
30 60 90 120 150
Problem Size
Results are mixed (no approach consistently better).
But differences are again relatively small.
Martin Pelikan, James D. Laury, Jr. Parallelization of Model Building in hBOA and Scalability
18. Results on hTrap
1e+06
hBOA
hBOA with ordered nodes
Number of Evaluations
100000
10000
1000
27 81 243
Problem Size
hBOA with node ordering in fact outperforms standard hBOA.
But differences are again relatively small.
Martin Pelikan, James D. Laury, Jr. Parallelization of Model Building in hBOA and Scalability
19. Results on 2D Ising Spin Glass
hBOA
hBOA with ordered nodes
Number of Evaluations
1000
100
36 64 100 144 196
Problem Size
hBOA with node ordering needs slightly more evaluations.
But differences are relatively small.
Martin Pelikan, James D. Laury, Jr. Parallelization of Model Building in hBOA and Scalability
20. Summary and Conclusions
Summary
Tested scalability of parallel model building in hBOA.
Analyzed the results.
Conclusions
Parallelization leads to a slight increase in the number of
function evaluations.
But the negative effects are negligible compared to the gains.
Bottom line: If model building is the bottleneck, parallelize
the model building!
Martin Pelikan, James D. Laury, Jr. Parallelization of Model Building in hBOA and Scalability
21. Acknowledgments
Acknowledgments
NSF; NSF CAREER grant ECS-0547013.
University of Missouri; High Performance Computing
Collaboratory sponsored by Information Technology Services;
Research Award; Research Board.
Martin Pelikan, James D. Laury, Jr. Parallelization of Model Building in hBOA and Scalability