approximate likelihood

[cranmer]

This is more of a note since merging notebooks is difficult.

I'm prototyping a simple likelihood-free inference stage instead of the known likelihood.

def lnlike(theta, x, phi):
    #exact likelihood for testing
    #mean, std = theta, 2 + np.sin(phi)
    #return np.log(norm(mean, std).pdf(x).prod())
    
    #approx likelihood
    simdata = simulator(theta, phi,n_samples=1000)
    # this xrange is specific to this simulator
    # better to have auto-range, but would need density to correct for range & bin width
    xrange = (3-9,3+9) 
    approx_density = Distribution(name='approx density', range=xrange,samples=simdata)
    return np.sum(np.log(approx_density.pdf(x)))

[cranmer]

PS> it makes it go slower :-)