Hyperparameter selection

In a machine learning environment having a plurality of hyperparameter configurations, each hyperparameter configuration including a set of hyperparameters applicable to a machine learning model, a method for identifying a hyperparameter configuration for deployment, comprising:

• generating a plurality of hyperparameter configurations applicable to the model;
• for each hyperparameter configuration of the plurality of hyperparameter configurations,
• iteratively evaluating a performance and a reward distribution indicating whether the evaluated performance will change during successive iterations, and
• indicating when a potential for improvement from additional iterations has subsided, based on:
• introducing a procedure based on a random number to update the reward distribution with the performance of the hyperparameter configuration on a given iteration indicating a likelihood of a change to the performance of the hyperparameter configuration as a result of successive iterations;
• and
• carrying out the procedure at each iteration;
• determining, at each iteration, if the performance metric is greater than a random value, the random value drawn from a distribution of possible performance metrics, and if so,
• adding the number of current iterations to an accumulation of a first distribution parameter, and if not,
• adding the number of current iterations to an accumulation of a second distribution parameter;
• ranking the hyperparameter configurations based on the reward randomly drawn from the distribution evaluated for each hyperparameter configuration;
• selecting at least one of the best performing hyperparameter configurations based on the ranking of hyperparameter configurations;
• terminating and discarding lower ranking hyperparameter configurations upon determining a convergence indicative of accurate performance and that the potential for improvement from additional iterations has subsided.

A computer program embodying program code on a non-transitory medium that, when executed by a processor, performs steps for implementing a method of selecting hyperparameters for improving performance of a machine learning (ML) model, the method comprising:

• receiving a plurality of hyperparameter configurations, each hyperparameter configuration of the plurality of hyperparameter configurations having a set of hyperparameters applicable to a model and a set of training data;
• computing, for each hyperparameter configuration, a performance and a corresponding reward distribution indicating whether the performance is converged to indicate an accuracy of the actual performance of the respective hyperparameter configuration;
• storing, in a data structure corresponding to each hyperparameter configuration, a representation of the performance and the reward distribution;
• ranking the hyperparameter configurations based on a reward value randomly sampled from the reward distribution;
• further comprising:
• for each hyperparameter configuration,
• iteratively applying the set of training data using the hyperparameter configuration;
• at each iteration, computing the performance;
• at each iteration, generating a random value within a range defined by the performance;
• comparing the random value to the computed performance, and
• if the computed performance is greater than the random value,
• adding a value equal to the number of iterations to a success counter, and if not adding the value equal to the number of iterations to a failure counter;
• generating a distribution of the reward for each hyperparameter configuration, with the shape of the distribution denoting that the computed performance accurately denotes an actual performance, further comprising:
• computing, for each hyperparameter configuration of the plurality of hyperparameter configurations, a distribution of the reward, the distribution indicative of whether the performance of the hyperparameter configuration is converged, the distribution representable as a graph denoting, for each hyperparameter configuration, a convergence of performance denoted by a peak at a location on the graph;
• terminating and discarding lower ranking hyperparameter configurations upon determining a convergence indicative of accurate performance and that a potential for improvement from additional iterations has subsided;
• and
• selecting at least one of the best performing hyperparameter configurations based on the ranking of hyperparameter configurations.