This is the not-so-secret sauce when training neural nets and backtesting for algorithmic trading. It dramatically reduces the time taken.
Genetic algorithms for training deep neural networks (2017)
11–20 of 23 posts
Re: Genetic algorithms for training deep neural networks (2017)
#12I lost it at "emerging revolution". My PhD thesis in 2000 already used genetic algorithms for seeds and it was hardly new then.
Your missing the point I think. GAs aren't new, same as neural networks which were thought up half a century ago. It's the finding that GAs work at such a scale, with 100 of layers and millions of nodes and even outperform SGD, that is of interest here.
Re: Genetic algorithms for training deep neural networks (2017)
#13The article is from last year but it's still extremely valuable and interesting. Exploring this topic is currently my primary hobby. Specifically, I've been using OpenAI's retro (Sonic, Contra, Mario, Donkey Kong and, more recently FZero) and comparing the ancient NEAT with more fashionable stuff like DQN, PPO, A3C and DDPG. With my extremely limited experience, NEAT seems to outperform all of these other algorithms.…
You are evolving the topology, but using regular gradient descent/backprop for any given network, correct?
Re: Genetic algorithms for training deep neural networks (2017)
#14The article is from last year but it's still extremely valuable and interesting. Exploring this topic is currently my primary hobby. Specifically, I've been using OpenAI's retro (Sonic, Contra, Mario, Donkey Kong and, more recently FZero) and comparing the ancient NEAT with more fashionable stuff like DQN, PPO, A3C and DDPG. With my extremely limited experience, NEAT seems to outperform all of these other algorithms.…
You are evolving the topology, but using regular gradient descent/backprop for any given network, correct?
No, in NEAT both the weights and topology are evolved. It is totally gradient-free.
Re: Genetic algorithms for training deep neural networks (2017)
#15I lost it at "emerging revolution". My PhD thesis in 2000 already used genetic algorithms for seeds and it was hardly new then.
Genetic algorithms are (and were already back in 2000) a pretty decent and -more importantly- generic solution to the problem of global optimization (as opposed to local optimization) when the problem to optimize has some sort of (maybe not-so-smooth) structure. Many of the recent "AI" development very often boil down to finding a local extremum using some sort of ski down the slope optimization program (aka "trainin…
How are genetic algorithms guaranteed in any way to give you a global optimum? I thought it might just work better when your search space is non smooth.
Re: Genetic algorithms for training deep neural networks (2017)
#16This is the not-so-secret sauce when training neural nets and backtesting for algorithmic trading. It dramatically reduces the time taken.
Sounds interesting, are you able to go into more detail?
Re: Genetic algorithms for training deep neural networks (2017)
#17I lost it at "emerging revolution". My PhD thesis in 2000 already used genetic algorithms for seeds and it was hardly new then.
Genetic algorithms are (and were already back in 2000) a pretty decent and -more importantly- generic solution to the problem of global optimization (as opposed to local optimization) when the problem to optimize has some sort of (maybe not-so-smooth) structure. Many of the recent "AI" development very often boil down to finding a local extremum using some sort of ski down the slope optimization program (aka "trainin…
Agreed. However, there is no reason to believe that a GA can give a better result than just randomly picking points in search space and doing local optimization around those points. GA's are evo-babble with much hand waving about 'correlation' during mutation or crossover rubbish.
The 'curse of dimensionality' in global optimization doesn't go away by using GA's nor can you talk about their effectiveness without characterizing the landscape you are searching (the no free lunch theorem).
Re: Genetic algorithms for training deep neural networks (2017)
#18I lost it at "emerging revolution". My PhD thesis in 2000 already used genetic algorithms for seeds and it was hardly new then.
Genetic algorithms are (and were already back in 2000) a pretty decent and -more importantly- generic solution to the problem of global optimization (as opposed to local optimization) when the problem to optimize has some sort of (maybe not-so-smooth) structure. Many of the recent "AI" development very often boil down to finding a local extremum using some sort of ski down the slope optimization program (aka "trainin…
Re: Genetic algorithms for training deep neural networks (2017)
#19Too many ideas, too little time! I am already thinking for a while about how Deep Learning and Genetic Algorithms could benefit each other.
GAs allow optimization of parameters without a differentiable loss function - a major problem with evaluating behavior of a neural model, for example.
But also GAs could benefit from ML/DL. Predicting loss functions from a chromosome representation (to save computing time), learning to select promising pairs and even learning cross-over operators.
Re: Genetic algorithms for training deep neural networks (2017)
#20Earlier quoted context omitted.
Genetic algorithms are (and were already back in 2000) a pretty decent and -more importantly- generic solution to the problem of global optimization (as opposed to local optimization) when the problem to optimize has some sort of (maybe not-so-smooth) structure. Many of the recent "AI" development very often boil down to finding a local extremum using some sort of ski down the slope optimization program (aka "trainin…
How are genetic algorithms guaranteed in any way to give you a global optimum? I thought it might just work better when your search space is non smooth.
They're not, as far as I know. In fact, that was one of the big selling points of reinforcement learning - that it tends to reach a better minimum than GA.