ChrisRackauckas
commented
4 years ago You hit an AD bug. I think it might be easier to just use the out-of-place (u,p,t) style instead of mutating (you had that syntax wrong anyways). With that, here's a workaround:
using DifferentialEquations, DiffEqFlux
using Optim
using LinearAlgebra
# --> Parameters of the cart.
m1 = 0.9
k1 = 0.1
# --> Parameters of the pendulum.
m2 = 0.1
k2 = 0.1
l = 10
g = 9.818
# --> Pack parameters.
p = [m1, m2, k1, k2, l, g]
# --> Define controller.
K = FastChain(
FastDense(4, 4, tanh),
FastDense(4, 1)
)
W = initial_params(K)
# --> Dynamics of the cartplot.
function controlled_pendulum!(u, p, t)
# --> Unpack variables.
x, θ, dx, dθ = u
# --> Unpack physical parameters.
m₁, m₂, k₁, k₂, l, g = p[1:6]
# --> Unpack neural net parameters.
W = p[7:end]
# --> Left-hand side of the equations of motion.
du = [dx,dθ,-m₂*l*dθ^2*sin(θ) - k₁*dx + K(u, W)[1],g*sin(θ) - k₂*dθ]
# --> Mass matrix.
M = reshape([1,0,0,0,
0,1,0,0,
0,0,m₁ + m₂,-m₂*l*cos(θ),
0,0,-cos(θ),l],4,4)
# --> Invert mass matrix.
return M \ du
end
# --> Setup the ODE problem.
u₀ = Float32[0.0, 0.1, 0.0, 0.0]
params = Float32[p ; W]
tspan = (0.0f0, 10.0f0)
prob = ODEProblem(controlled_pendulum!, u₀, tspan; p=params)
# --> Predict function.
function predict(θ)
Array(concrete_solve(prob, Tsit5(), u₀, [p; θ], saveat=0.1))
end
# --> Loss and callback functions.
function loss(θ)
return norm(predict(θ))^2
end
l = loss(W)
cb = function(θ, l)
println(l)
return false
end
# --> Train the controller.
res = DiffEqFlux.sciml_train(loss, W, BFGS(), cb=cb)Note your model is unstable, so you will need to play around with it (maybe change the initial condition on the controlled variable or something). But that weird mass matrix syntax is the workaround for the bug, and I'll get that upstreamed and fixed.
Hi guys,
I have been able to run successfully the Lotka-Voltera example. I am now looking at a classical benchmark of optimal control, namely the inverted pendulum on a cart where my controller is given by a fairly simple neural net (just to give it a try). Unfortunately, whenever I try to train the model, I get a
StackOverflowErrorand I am not sure how to solve my problem.Below is my code.
Whenever I run this code, here is the error I get :
I am pretty new to Julia so any help is welcome :)
++ JC