Optimize TF layers weigth with Genetics Algorithms
Gradient descending agorithm are well known for optimizing the weights of neural networks and finding the solution to converge into an optimal solution.
Usually the derivative is applied to a single output value and then the decision is gradually improved through an improvement in the optimization function (or target).
The purpose of this mini-project is to demonstrate that we can optimize the weights of a net in TensorFlow (keras) using a basic genetic algorithm initiated from a random state.
Genetic algorithms are based on the genetic process of biologically living organisms. Where in a closed environment of X species only the strongest (or the best adapted) to the environment survive in a new species to evolve to a new specie with a mixted gens of the survivors.
Genes->Parameter to be optimized
Species->Each species contains a certain amount of genes all together offer a solution to a specific problem.
Population-->Population is a group of species all together perform a generation
Evolution->Evolution is the step that is generated between one generation and another, where there is a mixture of genes that generates new species and new species are born with a certain mutation.
There is a varied range of problems to which the application of genetic algorithms can be useful(like salesman problem) specially if we want to optimize the weights of a neural network. For practical purposes, it is especially useful when the train dataset could be diferent at each step specially if we want to teach a network to make full decisions based on past decisions(outputs) of the same network.
For example the network receives inputs from distance sensors to an obstacle and the network can make several decisions (breaking, turning left/rigth, accelerating etc.). Take in consideration! each observation in the train dataset is completely dynamic because at each observation the distances of the sensor will be different based on the past decisions of the network.
Usually quantitave traders download financial timeseries then apply a treatment to the data if necessary (normalization PCA etc...). Finally they train a network(python/R etc..) normally with the objective of reducing the error between the prediction and the real value and finally they take a decision(automatic or not) based on model single prediction value.
Please note that the main problem of quantitative traders when training a network with static observations for the financial markets are completely exposed to high volatility, which is why most quantitative traders lose they money. They not only need a predictive model of a value. They also need a complex and operative system that is able to manage the capital effectively.
Theoretically and as demonstrated by Google with the "Alpha Zero AI" the net can learn in a more efficient way (and sometimes surprisingly) IF LEARNING PROCES IS NOT CONDITIONED/LIMITED BY HUMAN KWNOLADGE.
This is with genetic algorithms we can train a network to make complex decisions such as (buy,sell, parcial close, hedges etc..) all based on the current state of command controls of the trading platform and past net desicions. (FOR MORE INFORMATION ABOUT FIANACIAL STOCK WITH GENETICS ALGORHITMS PLEASE READ Project https://github.com/nopaixx/MaketsGA)
For easy implementation code we have based ourselves on the kaggle competition TAG SALT IDENTIFICATION (https://www.kaggle.com/c/tgs-salt-identification-challenge)
Pay attention to:
-def parallel_scoring ->Perform a custom fitness funcion higher value best
-def runtournament ->Perform a random natural seleccion of two best species to join
-def model_crossover ->Join best two species in random way and apply a muttation
-Check not model.fit() transaction is executed at any time
-As main objective of the document we have been able to optimize a neural network (in the case of the example Convultional) without applying the method of Gradient descence.
-We have set the basis for developing more complex loss functions without the use of tensors
-We have talked about how we can use genetic algorithms to train neural networks and how we can use neural networks from a different point of view.
-No doubt about the use of GA in neural networks presents an advantage for certain types of problem in which for example the traing set observations are dynamic and varied in each training episode
-This can be the beginning of a more complex and optimized library to develop advanced quantitative trading models where the predictive model does not consist of a buy or sell signal in itself, but the model consists of a complex trading system in itself.
-I think that this may constitute the beginning of a promosing and completely self-adaptable trading system and discover automated trading systems more efficient and not limited by human knowledge.