Modified Brain Storm Optimization
David Kuo
Osborn’s Original Rule
Suspend Judgment
There is no idea as bad idea.
1
Anything goes
Anything coming to your mind during process is an idea.
2
Cross-fertilize
Lots of ideas can and should be based on existed ideas.
3
Go for Quantity
It is necessary to generate as many ideas as possible.
4
Brain Storm Process
Generate ideas
(generate operation)
Cluster ideas
(cluster operation)
Select & Merge
ideas
Evaluate
cons & pros
Get good ideas
Hopefully to obtain a good
solution
Brain Storm Algorithm
1. Generate n potential solution (individuals)
2. Cluster n individuals into k clusters using k-means algorithm.
3. Evaluate the n individuals.
4. Rank individuals in each other cluster and record the best one as cluster center.
5. Select a cluster center or generate a individual to replace it (P5).
6. Generate new individuals from one or more (P6b) clusters:
a) Choose one cluster (P6ai) and choose cluster center or not (P6biii).
b) Choose more than one clusters and choose cluster center or not (P6c).
7. If n new individuals have been generated, go to step8, otherwise go back to step6.
8. Terminate if reached target, otherwise go back step2.
Generate new individual
Formula:
𝑋 𝑛𝑒𝑤 = 𝑋𝑠𝑒𝑙𝑒𝑐𝑡 +𝜉 ∗ 𝑛(𝜇, 𝜎)
𝜉 = 𝑙𝑜𝑔𝑠𝑖𝑔( 0.5 ∗ max_iteration − current_iternation) 𝑘 𝑥𝑖 ∗ 𝑟𝑎𝑛𝑑()
Weight
Gaussian Function
Logsig(n) is “Logarithmic sigmoid transfer function”
Experiments
Experiments
Modified Brain Strom Optimization
Using Simple Grouping replaced k-means1
Idea difference Strategy2
Simple Grouping Method
1. Randomly select M different ideas from current generation as the seeds of the M
groups and denoted as S𝑗 (1 ≤ j ≤ M)
2. For each ideas 𝑋𝑖(1 ≤ 𝑖 ≤ 𝑁) in the current generation, calculate its distance to each
group j as:
3. Compare all the M distance values and assign the idea 𝑋𝑖 into the nearest group,
but 𝑆𝑗unchanged.
4. Go to step2 for the next idea, otherwise all ideas have been assigned.
𝑑𝑖𝑗 = 𝑋𝑖, 𝑆𝑗 = (
𝑑=1
𝐷
(𝑋𝑖𝑑 − 𝑆𝑖𝑑)2)/𝐷
Idea Different Strategy
𝜉 𝑜𝑟𝑖𝑔𝑖𝑛 = 𝑙𝑜𝑔𝑠𝑖𝑔( 0.5 ∗ max_iteration − current_iternation) 𝑘 𝑥𝑖 ∗ 𝑟𝑎𝑛𝑑()
• In every generation, we use this FIXED value function.
No any feedback information from search process, we may not catch the search
characteristics.
• Commonly intuition
Noise is larger in early evolutionary phase and gradually become smaller during running
by the control of the logsig function.
Idea Different Strategy (cond)
𝜉 𝑜𝑟𝑖𝑔𝑖𝑛 = 𝑙𝑜𝑔𝑠𝑖𝑔( 0.5 ∗ max_iteration − current_iternation) 𝑘 𝑥𝑖 ∗ 𝑟𝑎𝑛𝑑()
• Range
1. Logsig function return value (0,1) and ξ is also in (0,1).
2. The value ξ multiplies Gaussian random value(μ=0,σ=1) is with high probability within
the range (-4,4).
𝑋 𝑛𝑒𝑤 = 𝑋𝑠𝑒𝑙𝑒𝑐𝑡 +𝜉 ∗ 𝑛(𝜇, 𝜎)
Idea Different Strategy (cond)
How to improve this operator?
Idea Different Strategy (cond)
𝑋 𝑛𝑒𝑤_d =
𝑟𝑎𝑛𝑑𝑜𝑚 𝐿 𝑑, 𝐻 𝑑 , 𝑖𝑓 𝑟𝑎𝑛𝑑𝑜𝑚 0,1 < 𝑝 𝑟
𝑋 𝑜𝑟𝑖_𝑑 + 𝑟𝑎𝑛𝑑𝑜𝑚 0,1 ∗ 𝑋 𝑎𝑑 − 𝑋 𝑏𝑑 , 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
Discuss from two distinct ideas from all current
ideas.
Based on existed idea.
Find a new idea. Control new ideas created.
• Parameters setting
1. Experiments indicate that the value of 𝑝 𝑟 within [0.001,0.01] would be good.
2. In this paper, we choose 𝑝 𝑟=0.005.
Experimental Studies
300000 generations
30 times for each function.
Algorithms configurations
Experimental Studies
• MBSO is better than BSO.
This may be due to IDS can match the environment to provide
suitable noise.
Experimental Studies
Experimental Studies
• How to set 𝑝 𝑟?
a) Unimodal function
1. To observe this figure, we can find 𝑝 𝑟 is smaller, solution accuracy is more higher.
b) Multimodal function
1. Too much lower or higher is harmful for algorithm performance.
2. The results indicate that a 𝑝 𝑟 within [0.001,0.01].
Conclusion
Simple Grouping Method
• Advantages:
1. Simple to implement.
2. Lower computational burden.
• Disadvantages:
1. Can’t identify different peeks in multimodal function.
Idea Different Strategy
• Advantages:
1. Lower computation burden.
2. Should match environment of the evolutionary process.
• Disadvantages:
1. Can’t identify different peeks in multimodal function.
• MBSO using SGM and IDS has higher performance than BSO.
Question?

Brain storm algorithm

  • 1.
    Modified Brain StormOptimization David Kuo
  • 2.
    Osborn’s Original Rule SuspendJudgment There is no idea as bad idea. 1 Anything goes Anything coming to your mind during process is an idea. 2 Cross-fertilize Lots of ideas can and should be based on existed ideas. 3 Go for Quantity It is necessary to generate as many ideas as possible. 4
  • 3.
    Brain Storm Process Generateideas (generate operation) Cluster ideas (cluster operation) Select & Merge ideas Evaluate cons & pros Get good ideas Hopefully to obtain a good solution
  • 4.
    Brain Storm Algorithm 1.Generate n potential solution (individuals) 2. Cluster n individuals into k clusters using k-means algorithm. 3. Evaluate the n individuals. 4. Rank individuals in each other cluster and record the best one as cluster center. 5. Select a cluster center or generate a individual to replace it (P5). 6. Generate new individuals from one or more (P6b) clusters: a) Choose one cluster (P6ai) and choose cluster center or not (P6biii). b) Choose more than one clusters and choose cluster center or not (P6c). 7. If n new individuals have been generated, go to step8, otherwise go back to step6. 8. Terminate if reached target, otherwise go back step2.
  • 5.
    Generate new individual Formula: 𝑋𝑛𝑒𝑤 = 𝑋𝑠𝑒𝑙𝑒𝑐𝑡 +𝜉 ∗ 𝑛(𝜇, 𝜎) 𝜉 = 𝑙𝑜𝑔𝑠𝑖𝑔( 0.5 ∗ max_iteration − current_iternation) 𝑘 𝑥𝑖 ∗ 𝑟𝑎𝑛𝑑() Weight Gaussian Function Logsig(n) is “Logarithmic sigmoid transfer function”
  • 6.
  • 7.
  • 8.
    Modified Brain StromOptimization Using Simple Grouping replaced k-means1 Idea difference Strategy2
  • 9.
    Simple Grouping Method 1.Randomly select M different ideas from current generation as the seeds of the M groups and denoted as S𝑗 (1 ≤ j ≤ M) 2. For each ideas 𝑋𝑖(1 ≤ 𝑖 ≤ 𝑁) in the current generation, calculate its distance to each group j as: 3. Compare all the M distance values and assign the idea 𝑋𝑖 into the nearest group, but 𝑆𝑗unchanged. 4. Go to step2 for the next idea, otherwise all ideas have been assigned. 𝑑𝑖𝑗 = 𝑋𝑖, 𝑆𝑗 = ( 𝑑=1 𝐷 (𝑋𝑖𝑑 − 𝑆𝑖𝑑)2)/𝐷
  • 10.
    Idea Different Strategy 𝜉𝑜𝑟𝑖𝑔𝑖𝑛 = 𝑙𝑜𝑔𝑠𝑖𝑔( 0.5 ∗ max_iteration − current_iternation) 𝑘 𝑥𝑖 ∗ 𝑟𝑎𝑛𝑑() • In every generation, we use this FIXED value function. No any feedback information from search process, we may not catch the search characteristics. • Commonly intuition Noise is larger in early evolutionary phase and gradually become smaller during running by the control of the logsig function.
  • 11.
    Idea Different Strategy(cond) 𝜉 𝑜𝑟𝑖𝑔𝑖𝑛 = 𝑙𝑜𝑔𝑠𝑖𝑔( 0.5 ∗ max_iteration − current_iternation) 𝑘 𝑥𝑖 ∗ 𝑟𝑎𝑛𝑑() • Range 1. Logsig function return value (0,1) and ξ is also in (0,1). 2. The value ξ multiplies Gaussian random value(μ=0,σ=1) is with high probability within the range (-4,4). 𝑋 𝑛𝑒𝑤 = 𝑋𝑠𝑒𝑙𝑒𝑐𝑡 +𝜉 ∗ 𝑛(𝜇, 𝜎)
  • 12.
    Idea Different Strategy(cond) How to improve this operator?
  • 13.
    Idea Different Strategy(cond) 𝑋 𝑛𝑒𝑤_d = 𝑟𝑎𝑛𝑑𝑜𝑚 𝐿 𝑑, 𝐻 𝑑 , 𝑖𝑓 𝑟𝑎𝑛𝑑𝑜𝑚 0,1 < 𝑝 𝑟 𝑋 𝑜𝑟𝑖_𝑑 + 𝑟𝑎𝑛𝑑𝑜𝑚 0,1 ∗ 𝑋 𝑎𝑑 − 𝑋 𝑏𝑑 , 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 Discuss from two distinct ideas from all current ideas. Based on existed idea. Find a new idea. Control new ideas created. • Parameters setting 1. Experiments indicate that the value of 𝑝 𝑟 within [0.001,0.01] would be good. 2. In this paper, we choose 𝑝 𝑟=0.005.
  • 14.
    Experimental Studies 300000 generations 30times for each function. Algorithms configurations
  • 15.
    Experimental Studies • MBSOis better than BSO. This may be due to IDS can match the environment to provide suitable noise.
  • 16.
  • 17.
    Experimental Studies • Howto set 𝑝 𝑟? a) Unimodal function 1. To observe this figure, we can find 𝑝 𝑟 is smaller, solution accuracy is more higher. b) Multimodal function 1. Too much lower or higher is harmful for algorithm performance. 2. The results indicate that a 𝑝 𝑟 within [0.001,0.01].
  • 18.
    Conclusion Simple Grouping Method •Advantages: 1. Simple to implement. 2. Lower computational burden. • Disadvantages: 1. Can’t identify different peeks in multimodal function. Idea Different Strategy • Advantages: 1. Lower computation burden. 2. Should match environment of the evolutionary process. • Disadvantages: 1. Can’t identify different peeks in multimodal function. • MBSO using SGM and IDS has higher performance than BSO.
  • 19.