Solving Random Ordinary Differential Equations

In this tutorial we will solve a RODE with NNRODE. Consider the equation:

\[du = f(u,p,t,W)dt\]

where $f(u,p,t,W)=2u\sin(W)$ and $W(t)$ is a Noise process.

f = (u,p,t,W) ->   2u*sin(W)
tspan = (0.00f0, 1.00f0)
u0 = 1.0f0
dt = 1/20f0

We start off by defining the NoiseProcess $W(t)$. In this case, we define a simple Gaussian Process. See Noise Processes for defining other types of processes.

W = WienerProcess(0.0,0.0,nothing)

Then, we need to define our model. In order to define a model, we can use Flux.chain or DiffEqFlux.FastChain.

chain = Flux.Chain(Dense(2,5,elu),Dense(5,1)) #Model using Flux
chain = FastChain(FastDense(2,50,tanh), FastDense(50,2)) #Model using DiffEqFlux

And let's define our optimizer function:

opt = ADAM(1e-3)

Now, let's pass all the parameters to the algorithm and then call the solver. If we already have some initial parameters, we can pass them into the NNRODE as well.

alg = NNRODE(chain , W , opt , init_params)
sol = solve(prob, NeuralPDE.NNRODE(chain,W,opt), dt=dt, verbose = true,
            abstol=1e-10, maxiters = 15000)