1. Non-revisiting Genetic Algorithm with Constant Memory
Yang LOU, Shiu Yin YUEN
Department of Electronic Engineering
City University of Hong Kong

2. Contents Introduction of cNrGA Two Pruning Strategies
Experimental Results
Conclusions

3. Abbreviations NrGA – Non-revisiting Genetic Algorithm
cNrGA – NrGA for continuous problems
BSP Tree – Binary Space Partitioning Tree (Archive Structure in (c-)NrGA)
LRU – Least Recently Used
R – Random
MT – Memory Threshold

4. Brief Intro. to cNrGA
Re-generate the same individual – uniform crossover;
Parameter-less adaptive mutation performed within sub-regions;
Re-evaluation is avoided.

5. cNrGA
Use the entire search history and parameter- less adaptive mutation
Usage of memory is directly proportional to the Max number of Function Evaluations (MemoryUsage ∝ MaxFEs)
Conventionally, MaxFEs = 40,000 is recommended for problems such as CEC test suite
To Keep
To Overcome
To Overcome

6. BSP Tree and Sub-regions
2D Example
[Upper Pic] Insertion of nodes (solutions) a and b
[Lower Pic] Insertion of node (solution) c
A node = {solution, fitness, time stamp, info. of sub-region, etc.}

7. Contents Introduction of cNrGA Two Pruning Strategies
Experimental Results
Conclusions

8. Strategies to Prune BSP Tree
Basic Idea: When the memory threshold (MT) is reached, an element in the archive is pruned before recording a new one.
Least Recently Used (LRU) Pruning
To delete the most useless information in the archive, time stamp is necessary
Random (R) Pruning
To randomly delete some information in the archive

9. Comparison of LRU &. R
LRU &. R: Remove two nodes (mutual siblings) from the BST tree, and add back the nodes recording the latest generated solutions
LRU vs. R: Choose different history information to delete

10. Technically suggest not to prune node a
Technically easy to prune node b
Re-allocate the sub- regions for related nodes
Delete historical information of node b

11. The Novel Operators cNrGA/CM/LRU &. cNrGA/CM/R
LRU &. R: Novel Parameter-less Adaptive Mutation Operators
Introduce no additional parameters
Enlarge the mutation range
LRU: Non-uniform - Older solutions are prone deleted and their search sub-regions are expanded
R: Uniform
0.5
0.25
(Exploration ) *

12. Contents Introduction of cNrGA Two Pruning Strategies
Experimental Results
Conclusions

13. Experimental Study Compare: cNrGA, cNrGA/CM/LRU, and cNrGA/CM/R
CEC 2013 benchmark functions
Parameters:
Population Size: 100
Crossover Rate: 0.5
Very few parameters in NrGA

14. Experimental Study 𝐼𝑛𝑓𝑜.𝐿𝑜𝑠𝑠= 𝑀𝑎𝑥𝐹𝐸𝑠−𝑀𝑇 𝑀𝑎𝑥𝐹𝐸𝑠 MaxFEs MT Info. Loss
𝐼𝑛𝑓𝑜.𝐿𝑜𝑠𝑠= 𝑀𝑎𝑥𝐹𝐸𝑠−𝑀𝑇 𝑀𝑎𝑥𝐹𝐸𝑠
MaxFEs
40,000
50,000
60,000
100,000 MT
30,000
10,000
Info.
 Loss
25%
40%
50%
90%

15. Average Rank (i) 25% Loss MaxFEs = 40,000 MT = 30,000 cNrGA
cNrGA/
CM/LRU
cNrGA/CM/R
Rank
1.96
1.86
2.18

16. Average Rank (ii) 40% Loss MaxFEs = 50,000 MT = 30,000 cNrGA
cNrGA/
CM/LRU
cNrGA/CM/R
Rank 2
2.11
1.89

17. Average Rank (iii) 50% Loss MaxFEs = 60,000 MT = 30,000 cNrGA
cNrGA/
CM/LRU
cNrGA/CM/R
Rank
1.93 2
2.07

18. Average Rank (iv) 90% Loss MaxFEs = 100,000 MT = 10,000 cNrGA
cNrGA/
CM/LRU
cNrGA/CM/R
Rank 2
1.93
1.65

19. Contents Introduction of cNrGA Two Pruning Strategies
Experimental Results
Conclusions

20. Conclusions
The applicability of cNrGA is widen to the situations when MaxFEs is substantial;
Least Recently Used (LRU) and Random (R) pruning strategies are proposed to maintain the memory usage constant ;
Both pruning strategies maintain the good performance of cNrGA.

21. Thanks for your attention!
Q & A