MADGRAD

This implements MADGRAD (A M omentumized, Ad aptive, Dual Averaged Grad ient Method for Stochastic Optimization) for stochastic optimisation as published in https://arxiv.org/abs/2101.11075.

Usage

e["Solver"]["Type"] = "Optimizer/MADGRAD"

Results

These are the results produced by this solver:

Best Gradient(x)
  • Usage: e[“Results”][“Best Gradient(x)”] = List of real number

  • Description: Values of dF(x) for the x parameters that produced the best F(x) found so far.

Best F(x)
  • Usage: e[“Results”][“Best F(x)”] = real number

  • Description: Optimal value of F(x) found so far.

Best Parameters
  • Usage: e[“Results”][“Best Parameters”] = List of real number

  • Description: Value for the x parameters that produced the best F(x).

Variable-Specific Settings

These are settings required by this module that are added to each of the experiment’s variables when this module is selected.

Lower Bound
  • Usage: e[“Variables”][index][“Lower Bound”] = real number

  • Description: [Hint] Lower bound for the variable’s value.

Upper Bound
  • Usage: e[“Variables”][index][“Upper Bound”] = real number

  • Description: [Hint] Upper bound for the variable’s value.

Initial Value
  • Usage: e[“Variables”][index][“Initial Value”] = real number

  • Description: [Hint] Initial value at or around which the algorithm shall start looking for an optimum.

Initial Mean
  • Usage: e[“Variables”][index][“Initial Mean”] = real number

  • Description: [Hint] Initial mean for the proposal distribution. This value must be defined between the variable’s Mininum and Maximum settings (by default, this value is given by the center of the variable domain).

Initial Standard Deviation
  • Usage: e[“Variables”][index][“Initial Standard Deviation”] = real number

  • Description: [Hint] Initial standard deviation of the proposal distribution for a variable (by default, this value is given by 30% of the variable domain width).

Minimum Standard Deviation Update
  • Usage: e[“Variables”][index][“Minimum Standard Deviation Update”] = real number

  • Description: [Hint] Lower bound for the standard deviation updates of the proposal distribution for a variable. Korali increases the scaling factor sigma if this value is undershot.

Values
  • Usage: e[“Variables”][index][“Values”] = List of real number

  • Description: [Hint] Locations to evaluate the Objective Function.

Configuration

These are settings required by this module.

Eta
  • Usage: e[“Solver”][“Eta”] = real number

  • Description: Learning Rate (Step Size)

Weight Decay
  • Usage: e[“Solver”][“Weight Decay”] = real number

  • Description: Smoothing factor for variable update.

Epsilon
  • Usage: e[“Solver”][“Epsilon”] = real number

  • Description: Term to facilitate numerical stability

Termination Criteria

These are the customizable criteria that indicates whether the solver should continue or finish execution. Korali will stop when at least one of these conditions are met. The criteria is expressed in C++ since it is compiled and evaluated as seen here in the engine.

Min Gradient Norm
  • Usage: e[“Solver”][“Min Gradient Norm”] = real number

  • Description: Specifies the minimal norm for the gradient of function with respect to Parameters.

  • Criteria: (_k->_currentGeneration > 1) && (_gradientNorm <= _minGradientNorm)

Max Gradient Norm
  • Usage: e[“Solver”][“Max Gradient Norm”] = real number

  • Description: Specifies the minimal norm for the gradient of function with respect to Parameters.

  • Criteria: (_k->_currentGeneration > 1) && (_gradientNorm >= _maxGradientNorm)

Max Value
  • Usage: e[“Solver”][“Max Value”] = real number

  • Description: Specifies the maximum target fitness to stop maximization.

  • Criteria: _k->_currentGeneration > 1 && (+_bestEverValue > _maxValue)

Min Value Difference Threshold
  • Usage: e[“Solver”][“Min Value Difference Threshold”] = real number

  • Description: Specifies the minimum fitness differential between two consecutive generations before stopping execution.

  • Criteria: _k->_currentGeneration > 1 && (fabs(_currentBestValue - _previousBestValue) < _minValueDifferenceThreshold)

Max Model Evaluations
  • Usage: e[“Solver”][“Max Model Evaluations”] = unsigned integer

  • Description: Specifies the maximum allowed evaluations of the computational model.

  • Criteria: _maxModelEvaluations <= _modelEvaluationCount

Max Generations
  • Usage: e[“Solver”][“Max Generations”] = unsigned integer

  • Description: Determines how many solver generations to run before stopping execution. Execution can be resumed at a later moment.

  • Criteria: _k->_currentGeneration > _maxGenerations

Default Configuration

These following configuration will be assigned by default. Any settings defined by the user will override the given settings specified in these defaults.

{
"Epsilon": 1e-06,
"Eta": 0.01,
"Model Evaluation Count": 0,
"Termination Criteria": {
    "Max Generations": 10000000000,
    "Max Gradient Norm": 1000000000000.0,
    "Max Model Evaluations": 1000000000,
    "Max Value": Infinity,
    "Min Gradient Norm": 1e-12,
    "Min Value Difference Threshold": -Infinity
    },
"Variable Count": 0,
"Weight Decay": 0.9
}

Variable Defaults

These following configuration will be assigned to each of the experiment variables by default. Any settings defined by the user will override the given settings specified in these defaults.

{
"Initial Mean": NaN,
"Initial Standard Deviation": NaN,
"Initial Value": NaN,
"Lower Bound": -Infinity,
"Minimum Standard Deviation Update": 0.0,
"Upper Bound": Infinity,
"Values": []
}