bsmart.scans.CMAES_ND

Optimisation using CMAES with Novelty Detection

Requires pip install cmaes, see: https://github.com/CyberAgentAILab/cmaes

BSMArt scan to be describe in an upcoming publication.

This scan uses cmaes to optimise a log-likelihood constructed from selected observables.

All desired observables must use the scaling CFUNCTION with a range of values for which the observable is in an acceptable range (other observables should have “SCALING”: “OFF” or at least no MEAN/VARIANCE/RANGE set so that they will be ignored.)

It is recommended to use this also with a validity condition, so that no further codes are evaluated when the observables are outside of this range, and the code chain can focus on first obtaining acceptable values for that/those observables.

For example, here we tell BSMArt to run SPheno and check the Higgs mass lies within 3 GeV of its observed value. We also record the lightest neutralino and selectron masses, construct a derived observable based on the difference of their masses, and insist that it is greater than zero:

"Codes": {
        "SPheno": {
            "Command": "< >",
            "InputFile": "< >",
            "OutputFile": "< >",
            "Observables":
            {
                        "mh": {
                        "SLHA": [
                                "MASS",
                                [25]
                        ],
                        "SCALING": "CFUNCTION",
                        "MEAN": 125.09,
                        "VARIANCE": 3,
                        "VALID": [122.09,128.09]
                        },
                        "mchi1": {
                                "SLHA" : [ "MASS", [1000022]], "SCALING": "OFF"]
                        },
                        "mse1": {
                                "SLHA" : [ "MASS", [1000011]], "SCALING": "OFF"]
                        },

                        "dmse1mchi1" : {
                        "FUNCTION": "mse1-mchi1",
                        "SCALING": "CFUNCTION",
                        "RANGE": [0,100000]
                        }
             },
             "Conditions": ["mse1>mchi1"],
             "Run": "True"
        }

Information

BSMArt Name: CMAES_ND

Requires:

• cmaes

• pyod

• sklearn

• traceback

• pandas

Settings:

• CMAESPopulationSize: Number of points to run per batch (optional, auto-calculated if not set)

• CMAESMean: Initial mean for CMA-ES optimizer (optional, random if not set)

• CMAESSigma: Initial step size for CMA-ES (optional, default: 1/sqrt(n_variables))

• CMAESMaxGenerations: Number of batches to run (default: 100)

• CMAES MaxLoss: Maximum loss value override (optional)

• CMAESPathSeeds: Path to seed file (.csv or .parquet) for initial mean (optional)

• CMAESNoveltyDetection: Enable novelty detection using HBOS (default: True)

• CMAESHBOSNBins: Number of bins for HBOS novelty detection (default: 100)

• CMAESFocus: Dict with Variables and Observables lists for novelty detection focus (optional)

• CMAESHierarchy: List of hierarchical observables to focus on first (optional)

• CMAESNoGoodPointsPatience: Number of generations without good points before stopping (default: 5000)

• CMAESGoodPointsEarlyStop: Target number of good points for early stopping (default: 5000)

• CMAESBestLossPatience: Patience for best loss counter (default: 200)

• CMAESBestNValidConstraintsPatience: Patience for best n valid constraints counter (default: 2000)

• Initial Sample: Path to initial sample file for warm start (optional)

class bsmart.scans.CMAES_ND.HBOSPenaliser(n_bins=100, *args, **kwargs)

Bases: object

fit(X)

get_penalties(x)

class bsmart.scans.CMAES_ND.NewScan(inputs, log)

Bases: Scan

Scanner class for CMAES-ND Scans

get_losses(observables)

Returns a list of losses.

The C-function loss should be zero if the observable is within allowed bounds, and greater than zero outside it.

If we include a validity condition, then other observables, evaluated afterwards, will be returned as NaN.

-> We should assign these the maximum loss ~ 710.

This is then compatible with the explicit use of ‘hierarchical observables’, where all other observables are set to the maximum.

initialise()

method to allow the user scan to overload run settings etc during the initialisation process

postprocess(Point, observables, data_point, temp_dir, log, lock=None)

Default postprocessing method for use with SLHA output, you should overload this for more interesting/sophisticated scans

run()

Algorithm:

Our initial samples are in [0,1] since we use scalers. Subsequent rounds use the normalizing flow to generate samples

class bsmart.scans.CMAES_ND.Optimiser(direction='min', *args, **kwargs)

Bases: object

fit(X)

get_penalties(X)