Lotka-Volterra with Flux.train!

The following is a quick example of optimizing Lotka-Volterra parameters using the Flux.jl style. As before, we define the ODE that we want to solve:

using DiffEqFlux, DiffEqSensitivity, Flux, OrdinaryDiffEq, Zygote, Test #using Plots

function lotka_volterra(du,u,p,t)
  x, y = u
  α, β, δ, γ = p
  du[1] = dx = (α - β*y)x
  du[2] = dy = (δ*x - γ)y
p = [2.2, 1.0, 2.0, 0.4]
u0 = [1.0,1.0]
prob = ODEProblem(lotka_volterra,u0,(0.0,10.0),p)

Then we define our loss function with solve for the adjoint method:

function predict_rd()
loss_rd() = sum(abs2,x-1 for x in predict_rd())

Note that the parameters p to be optimized are not passed as arguments to the loss_rd or predict_rd functions. This is because Flux.train! below uses implicit globals.

Now we setup the optimization. Here we choose the ADAM optimizer. To tell Flux what our parameters to optimize are, we use Flux.params(p). To make the optimizer run for 100 steps we use 100 outputs of blank data, i.e. Iterators.repeated((), 100) (this is where minibatching would go!).

opt = ADAM(0.1)
cb = function ()

# Display the ODE with the current parameter values.
Flux.train!(loss_rd, Flux.params(p), Iterators.repeated((), 100), opt, cb = cb)

And now p will be the optimal parameter values for our chosen loss function.