An introduction to Genetic Algorithms (Dublin Data Science)
Download as PPTX, PDF
0 likes296 views
Genetic algorithms are very simple, yet very powerful. In this talk I'll explain how they work and what they can do
Convert to study guideBETA
Transform any presentation into a summarized study guide, highlighting the most important points and key insights.
![while True:
stack = population.sort(key=fitness) # Sort by fitness
fittest = stack[n_cull:] # Delete the weakest
population = []
for i in range(n_cull):
# Spawn mutated copies
population.append(random.choice(fittest).mutate())
population.extend(fittest)
in 7 lines of python:](https://image.slidesharecdn.com/2018-06-24geneticalgorithms-180601080914/85/An-introduction-to-Genetic-Algorithms-Dublin-Data-Science-15-320.jpg)
An introduction to Genetic Algorithms (Dublin Data Science)
- 4. Failing to train a recommendation system Test accuracy Time 16%
- 5. Recommender Learned embedding Lots of hyper-parameters Embedding size Learning rate Number of deep layers etc....
- 6. Picking an embedding size 0.01 ? 0.03 ? 0.05 ? 1.0 ? 10.0 ? Help!!
- 7. So lets try a genetic algorithm!
- 10. argmaxx f(x)
- 11. x1 x2 x3 x4 x5 x6 x7 x8 We create a population...
- 12. 1. Remove the weakest 2. Spawn mutated copies Repeat many times!
- 13. This two-step algorithm is able to maximise f(x)
- 14. The population fans-out over time DeepMind: Population based training of neural networks
- 15. while True: stack = population.sort(key=fitness) # Sort by fitness fittest = stack[n_cull:] # Delete the weakest population = [] for i in range(n_cull): # Spawn mutated copies population.append(random.choice(fittest).mutate()) population.extend(fittest) in 7 lines of python:
- 16. Benefits of genetic algorithms
- 17. They can be applied to diverse problems Neural networks antenna Race car humans
- 18. They often come up with designs humans wouldnt have thought of
- 19. They can navigate large and complex state spaces
- 20. They dont require functions to be differentiable
- 22. They take a lot of time/compute 6 million years =
- 24. Appendix
- 25. Designing your genetic algorithm Asexual vs sexual breeding Breed top performers or all? How often to cull? Mutation rate
- 26. Bonus: Population based training Lets combine gradient descent with genetic algorithms for fun and profit Good efficiency trade-off between trying many options and not discarding work done diagram
- 30. Genetic algorithm vs Gradient descent Genetic algorithm Gradient descent Works for any function Function must be differentiable Slower Faster Requires you to design a random mutation function Good for generating answers to unknown spaces Good for generating answers to well modelled spaces Very easy to distribute Harder to distribute
- 31. They help escape your God complex