Improving the API, how to implement constraints and extended likelihood

[maxnoe]

Constraints

These are nothing special i suppose, we just need an operator to join likelihoods with different data sizes. I think the logical approach for the end user for a constraint would be something like this:

# unconstraint model for the data
x = var("x")
mu = par("mu")
sigma = par("sigma")
model = Normal(x,  mu, sigma)

# constraint to mu from other experiment:
mu_meas = var("mu_con")
sigma_meas = var("sigma_con")
constraint = Normal(mu_meas, mu, sigma_meas)

# join the models:
const_model = Join(model, constraint)

And then fit as usual

Extended Likelihood

It should be as simple as this from the user point of view:

x = var("x")
mu = par("mu")
sigma = par("sigma")
signal_model = Normal(x, mu, sigma)

tau = par("tau")
background_model = Exponential(tau)

N1 = par("N_sig")
N2 = par("N_bkg")

model = Join(N1, signal_model, N2, background_model)

Maybe we can implement the Extended likelihood over a parameter attribute? LIke N1 = par("N1", num_events=True) would trigger an extended Likelihood?

[ibab]

In my opinion, the most logical way to do it would be this: Require users to apply Join to reduce a vectorized likelihood to a scalar one (This would just sum the log-likelihoods in Theano). Then, use Join as well to combine two different models (sum their log-likelihood). A constraint would work like this:

# unconstrained model for the data
x = var("x")
mu = par("mu")
sigma = par("sigma")
model = Normal(x,  mu, sigma)

# constraint to mu from other experiment:
mu_meas = var("mu_con")
sigma_meas = var("sigma_con")
constraint = Normal(mu_meas, mu, sigma_meas)

# Take 
const_model = Join(Join(model), constraint)

Alternatively, the two different uses of Join could have different names, like All and Join or All and And. Using the same operator makes sense, though, as it does exactly the same thing (calculate a sum) in both cases. It would have to complain if passed a vector likelihood and a scalar likelihood at the same time. Multiple scalars or multiple vectors would be fine, though.

I don't like the idea of implementing special mechanisms for the Extended Likelihood, as it can be interpreted as a regular MLE model. What about implementing an operator Mix2E that can be used like this:

model = Mix2E(N1, model1, N2, model2)

This would probably have to return a scalar likelihood with the poisson constraint applied. Are there any cases where users might want to keep the model as a vector to modify it further? I can't think of any.

[ibab]

Oh, and when fit is called on a vector model, it would automatically be Joined so you wouldn't have to take this extra step for simple models.

[ibab]

Another idea: It should be possible to auto-join the model. The user would write something like this:

with Model() as model:
    # unconstrained model for the data
    x = var("x", observed=True)
    mu = var("mu")
    sigma = var("sigma")
    dist1 = Normal(x,  mu, sigma)

    # constraint to mu from other experiment:
    mu_meas = var("mu_con", observed=True)
    sigma_meas = var("sigma_con", observed=True)
    constraint = Normal(mu_meas, mu, sigma_meas)

(No explicit combination of dist and constraint is performed) Then model.fit would be used to get the best estimate for mu and sigma.

This is very similar to what's done in pymc.

[maxnoe]

I also thought, that one should maybe separate Model and Distribution. I do not know how one would make this work, but it looks good to me.

[KonstantinSchubert]

+1 on using an operator like Mix2E.

I like the use of Join() in both cases.

I agree that it should be possible to auto-join the model (passing a vector into the argument of a scalar PDF could trigger it, and fit could just receive a list of models which it then joins before fitting).

But I don't like magic. And I think a bit of verbosity on the user side helps catching mistakes and makes error messages more helpful to novices.