Variable parameter probability for machine-learning model generation and training

Abstract

A method includes generating, by a processor of a computing device, an output set of models corresponding to a first epoch of a genetic algorithm and based on an input set of models of the first epoch. The input set and the output set includes data representative of a neural network. The method includes determining a particular model of the output set based on a fitness function. A first topological parameter of a first model of the input set is modified to generate the particular model of the output set. The method includes modifying a probability that the first topological parameter is to be changed by a genetic operation during a second epoch of the genetic algorithm that is subsequent to the first epoch. The method includes generating a second output set of models corresponding to the second epoch and based on the output set and the modified probability.

What is claimed is: 
     
         1 . A computer system comprising:
 a memory configured to store instructions; and   a processor coupled to the memory and configured to execute the instructions to:
 set one or more probability values of an automated model building process based on a comparison of characteristics of two or more machine-learning models and a comparison of fitness metrics corresponding to the two or more machine-learning models; and 
 generate one or more output machine-learning models including executing the automated model building process using the one or more probability values such that a probability of the one or more output machine-learning models having a particular characteristic is based on the one or more probability values. 
   
     
     
         2 . The computer system of  claim 1 , wherein the particular characteristic includes at least one of a topological parameter, a layer parameter, or a node parameter. 
     
     
         3 . The computer system of  claim 1 , wherein the automated model building process is iterative, and wherein the two or more machine-learning models used to set the one or more probability values include a first machine-learning model from a first iteration of the automated model building process and a second machine-learning model from a second iteration of the automated model building process, wherein the second iteration is subsequent to the first iteration. 
     
     
         4 . The computer system of  claim 3 , wherein the second machine-learning model is based on the first machine-learning model. 
     
     
         5 . The computer system of  claim 3 , wherein executing the automated model building process includes performing one or more genetic operations to modify the first machine-learning model to generate the second machine-learning model, wherein modifications applied by the one or more genetic operations are based on probability values associated with the first iteration. 
     
     
         6 . The computer system of  claim 3 , wherein the first iteration and the second iteration are consecutive iterations. 
     
     
         7 . The computer system of  claim 1 , wherein the automated model building process is iterative, and wherein the two or more machine-learning models used to set the one or more probability values include a first machine-learning model from a first iteration of the automated model building process and a second machine-learning model from the first iteration of the automated model building process. 
     
     
         8 . The computer system of  claim 1 , wherein the automated model building process includes genetic operations and optimization operations. 
     
     
         9 . The computer system of  claim 1 , wherein the automatic model building process is configured to generate and evaluate machine learning models based on a specified input data set, and the fitness metrics are determined based on the specified input data set. 
     
     
         10 . A method comprising:
 setting, by a processor, one or more probability values of an automated model building process based on a comparison of characteristics of two or more machine-learning models and a comparison of fitness metrics corresponding to the two or more machine-learning models; and   generating, by the processor, one or more output machine-learning models including executing the automated model building process using the one or more probability values such that a probability of the one or more output machine-learning models having a particular characteristic is based on the one or more probability values.   
     
     
         11 . The method of  claim 10 , wherein the particular characteristic includes a number of nodes of the one or more output machine-learning models, a number of connections of the one or more output machine-learning models, a number of input nodes of the one or more output machine-learning models, a number of hidden layers of the one or more output machine-learning models, an activation function of at least one node of the one or more output machine-learning models, an aggregation function of at least one node of the one or more output machine-learning models, a bias function of at least one node of the one or more output machine-learning models, a layer type of one or more layers of the one or more output machine-learning models, or a combination thereof. 
     
     
         12 . The method of  claim 10 , further comprising providing the one or more output machine-learning models as input to a subsequent iteration of the automated model building process. 
     
     
         13 . The method of  claim 10 , wherein executing the automated model building process using the one or more probability values comprises determining a particular modification of the particular characteristic of the two or more machine-learning models to generate the one or more output machine-learning models, wherein the particular modification is determined based on the one or more probability values. 
     
     
         14 . The method of  claim 10 , further comprising:
 generating a first fitness value associated with a first machine-learning model of the two or more machine-learning models based on a fitness function; and   generating a second fitness value associated with a second machine-learning model of the two or more machine-learning models based on the fitness function, wherein the comparison of the fitness metrics corresponding to the two or more machine-learning models comprises a comparison of a difference between the second fitness value and the first fitness value to a threshold.   
     
     
         15 . The method of  claim 10 , further comprising identifying a beneficial difference between a first machine-learning model and a second machine-learning model based on the comparison of characteristics of the two or more machine-learning models and the comparison of fitness metrics corresponding to the two or more machine-learning models, wherein setting the one or more probability values includes increasing the probability that the one or more output machine-learning models include the beneficial difference. 
     
     
         16 . The method of  claim 10 , further comprising identifying a detrimental difference between a first machine-learning model and a second machine-learning model based on the comparison of characteristics of the two or more machine-learning models and the comparison of fitness metrics corresponding to the two or more machine-learning models, wherein setting the one or more probability values includes decreasing the probability that the one or more output machine-learning models include the detrimental difference. 
     
     
         17 . A computer-readable storage device storing instructions that, when executed, cause a computer to perform operations comprising:
 setting one or more probability values of an automated model building process based on a comparison of characteristics of two or more machine-learning models and a comparison of fitness metrics corresponding to the two or more machine-learning models; and   generating one or more output machine-learning models including executing the automated model building process using the one or more probability values such that a probability of the one or more output machine-learning models having a particular characteristic is based on the one or more probability values.   
     
     
         18 . The computer-readable storage device of  claim 17 , wherein executing the automated model building process comprises:
 generating a trainable model including modifying, based on the one or more probability values, at least one machine-learning model of the two or more machine-learning models;   sending the trainable model to a trainer;   receiving a trained model from the trainer; and   adding the trained model to an input set of models for a subsequent automated model building epoch, wherein the input set of models includes or corresponds to the one or more output machine-learning models.   
     
     
         19 . The computer-readable storage device of  claim 17 , wherein the automated model building process is iterative, and wherein the two or more machine-learning models used to set the one or more probability values include a first machine-learning model from a first iteration of the automated model building process and a second machine-learning model from a second iteration of the automated model building process, wherein the second iteration is subsequent to the first iteration. 
     
     
         20 . The computer-readable storage device of  claim 17 , wherein the automated model building process is iterative, and wherein the two or more machine-learning models used to set the one or more probability values include a first machine-learning model from a first iteration of the automated model building process and a second machine-learning model from the first iteration of the automated model building process.