Indentify parameters in an ODEProblem

[langestefan]

@cryptic.ax

Thanks for the feedback, we’re definitely open to improving documentation wherever possible. This is probably the page you were looking for, since it describes a similar kind of process to what you are doing here. I’ll see where I can put it so it’s more visible instead of being hidden under a folding menu.

Indeed, I was following that page already. But I got a totally lost with all the different options available to modify ODE problems.

@cryptic.ax

That looks correct

Apologies, I modified the code while you were typing a reply, i hope you caught that?

@cryptic.ax

The function you pass to GeneralLazyBufferCache takes the parameter values (as a vector) and does remake_buffer.

I implemented it like this:

# ordered symbols
syms = [sys.lC_i, sys.lC_e, sys.lC_h, sys.lR_ie, sys.lR_ea, sys.lR_ih, sys.lA_w, sys.lA_e]


@model function LGDS(x::AbstractArray, prob, params::Dict, t_int)

    # cache for remake(prob, p=newp)
    lbc = GeneralLazyBufferCache(function (p)
        remake_buffer(sys, prob.p, Dict(zip(syms, p)))    
    end)    

    # prior distributions
    σ ~ InverseGamma(2, 3)
    lC ~ MvNormal(params[:μ_C], params[:Σ_C])
    lR ~ MvNormal(params[:μ_R], params[:Σ_R])
    lA ~ MvNormal(params[:μ_A], params[:Σ_A])
   
    # remake the problem with the sampled parameter values and enforce order
    newp = lbc[[lC; lR; lA]]
    new_prob = remake(prob, p=newp)

    # solve the ODE
    T_ode = solve(new_prob, Tsit5(); save_idxs=1, verbose=false)
    if !(SciMLBase.successful_retcode(T_ode.retcode))
       return
    end

    # time points corresponding to observations `x`
    x_est = T_ode(time)  
    x ~ MvNormal(x_est.u, σ^2 * I)    

    return nothing
end

Is that appropriate?

I am bit skeptical of how much performance benefits it will actually yield, since each sample p would have to be cached and I don’t know enough about HMC or the NUTS sampler implementation to say whether they would be reused at all. Is there a way to check cache usage?