# Hyperparameters and its optimisation

## Hyperparameters and its optimisation

Hyperparameter is an important concept in domains like machine learning and deep learning. Hyperparameters are specifications or criterion values which helps in controlling the learning process and the specification of parameter requirements for initialising prior model training. The parameter values are determined with the help of training, generally, value means node weights.

Hyperparameters play a vital role in controlling the performance of the training algorithm. They also affect the working of training models. Hyperparameter impacts learning stage and its needed to be fluent and accurate in this stage to ensure proper working in real-time. Hyperparameters greatly influence the rate and quality of the learning phase of the model such as topology and its size, algorithm learning rate etc, Distinct model algorithm needs a number of parameters, few simple algorithms like regression require no parameters on the other and LASSO algorithm requires regularization hyperparameter which needs to be regulated prior to estimation of parameters through training algorithm.

Model parameters are the characteristics of training data which makes learning during training on its own with the help of machine learning like weights and split points in the decision tree. Hyperparameter is generally integer or discrete value which results in the mixed type optimisation. The occurrence of some of the parameter depends on the value of other factors, like the size of the middle layer depends on the number of layers present in the neutral network. Most of the time hyperparameters doesn't learn well in gradient-based methods used. These parameters can't be learnt by common optimisation by describing a model like error tolerance in SVM (simple vector machine).

Optimisation of hyperparameter is a process that results in an optimal model ensuring minimal loss function for provided data the objective function requires a pair of hyperparameters and gives corresponding loss. The method of determining most optimal hyperparameters in machine learning is known as hyperparameter optimisation. Some of the hyperparameter algorithms used are:

Grid search: Grid search makes the algorithm ready to face all combinations with the use of learning rate and the number of layers parameter. This helps in measuring the performance by cross-validation mechanism. This mechanism ministers for training model to get maximum patterns from the database. K Field cross-validation is a trustworthy validation method to get sufficient experience data for training the model. The grid model is easier and uses direct algorithm but faces problem in case data has high dimensional space.

Random Search: Random search method finds from spaces available and formulates sets with a probability distribution for random samples. The demerit of this method is, it doesn't use past data sets or experience to solve real-time problems and values are randomly picked from huge data sets.

Bayesian optimisation: The easiest way of optimisation is using the automatic model to approach, this is what called a bayesian optimisation algorithm. The basic model uses Gaussian Process for optimisation. It is used for approximation of objective function known as a surrogate model. After using Gaussian process for expressing the functional assumptions of optimisation, Acquisition function is used to build a utility function from posterior model to find the new sampling points for evaluation and maximisation. Hence Gaussian process exactly is the conversion to posterior function from prior distribution functions of data, done with the help of mathematical formulation called covariance matrix.

Some of the acquisition functions are:

MPI (maximum probability of improvement)

EI (Expected Improvement)

UCB (Upper Confidence Bound)

Another name given to hyperparameter tuning is Black Function. The result also depends on the grid search and data set chosen. If the correct combination of parameters is selected then the resulting algorithm will have higher accuracy and precision.