Specifying shape of parameter

[kjdoore]

I was wondering if it is possible to specify the shape of a parameter to be more than a single value. Here is an example where I specify three parameters that all have the same prior hyperparameters.

import numpy as np
import spotpy

class spot_setup(object):

    p1 = spotpy.parameter.Uniform(low=-10, high=10, optguess=1)
    p2 = spotpy.parameter.Uniform(low=-10, high=10, optguess=1)
    p3 = spotpy.parameter.Uniform(low=-10, high=10, optguess=1)

    def __init__(self, nobs=1000, range=[-5, 5], p=[1, 2, 3]):
        self.nobs = nobs
        self.p = p
        self.data_range = np.linspace(*range, nobs)
        self.obs = np.polyval(self.p, self.data_range)

    def simulation(self, x):
        parameters = [x['p1'], x['p2'], x['p3']]
        simulations = np.polyval(parameters, self.data_range)
        return simulations

    def evaluation(self):
        return self.obs

    def objectivefunction(self, simulation, evaluation):
        objectivefunction = spotpy.objectivefunctions.rmse(
            evaluation=evaluation, simulation=simulation
        )
        return objectivefunction

setup = spot_setup(nobs=50, p=[1, -5.2, 3])
sampler=spotpy.algorithms.sceua(setup, dbname='test_db', dbformat='csv')
sampler.sample(5000, ngs=20, kstop=3, peps=1e-10, pcento=1e-10)

What I am wanting is to be able to define one parameter p that has three values that are solved for. This is possible in PyMC with the shape keyword in the prior (e.g., p = pymc.Uniform("p", lower=-10, upper=10, shape=(3))) Is this something that can be done in spotpy? If not, is there any idea of how to do something similar?

My actual problem utilizes an arbitrary number of parameters that vary with different runs. So, I am wanting a way to define each parameter automatically rather than having to define multiple lines of pX = spotpy.parameter.Uniform(low=-10, high=10, optguess=1), where pX is the Xth parameter. Thanks!

[kjdoore]

After looking at this some more, it appears that switching the parameters away from class variables and into the __init__ will work. It does not make a single parameter p as I want, but it should suffice.

class spot_setup(object):

    def __init__(self, nobs=1000, data_range=[-5, 5], p=[1, 2, 3]):
        self.nobs = nobs
        self.p = p
        self.data_range = np.linspace(*data_range, nobs)
        self.obs = np.polyval(self.p, self.data_range)

        self.params = [spotpy.parameter.Uniform(name='p'+str(i), low=-10, high=10, optguess=0)
                       for i in range(len(self.p))]

    # This is needed since parameters are defined in the init. Otherwise, if class
    # variables (i.e., outside the init), they are generated automatically by the algorithm
    def parameters(self):
        return spotpy.parameter.generate(self.params)

    def simulation(self, x):
        parameters = [x['p'+str(i)] for i in range(len(self.p))]
        simulations = np.polyval(parameters, self.data_range)
        return simulations

    def evaluation(self):
        return self.obs

    def objectivefunction(self, simulation, evaluation):
        objectivefunction = spotpy.objectivefunctions.rmse(
            evaluation=evaluation, simulation=simulation
        )
        return objectivefunction

setup = spot_setup(nobs=50, p=[1, -5.2, 3])
sampler=spotpy.algorithms.sceua(setup, dbname='test_db', dbformat='csv')
sampler.sample(5000, ngs=20, kstop=3, peps=1e-10, pcento=1e-10)