Skip to content
Snippets Groups Projects
Commit e117c2da authored by Fredrik Bagge Carlson's avatar Fredrik Bagge Carlson
Browse files

updates

parent 4c0c7e32
No related branches found
No related tags found
No related merge requests found
Show whitespace changes
Inline Side-by-side
HRL/bibtexfile.bib 0 → 100644
+ 17
− 0
View file @ e117c2da
@article{DQN,
title={Human-level control through deep reinforcement learning},
author={Mnih, Volodymyr and Kavukcuoglu, Koray and Silver, David and Rusu, Andrei A and Veness, Joel and Bellemare, Marc G and Graves, Alex and Riedmiller, Martin and Fidjeland, Andreas K and Ostrovski, Georg and others},
journal={Nature},
volume={518},
number={7540},
pages={529--533},
year={2015},
publisher={Nature Research}
}
@article{lillicrap2015continuous,
title={Continuous control with deep reinforcement learning},
author={Lillicrap, Timothy P and Hunt, Jonathan J and Pritzel, Alexander and Heess, Nicolas and Erez, Tom and Tassa, Yuval and Silver, David and Wierstra, Daan},
journal={arXiv preprint arXiv:1509.02971},
year={2015}
}
TDAlambda/bibtexfile.bib 0 → 100644
+ 17
− 0
View file @ e117c2da
@article{DQN,
title={Human-level control through deep reinforcement learning},
author={Mnih, Volodymyr and Kavukcuoglu, Koray and Silver, David and Rusu, Andrei A and Veness, Joel and Bellemare, Marc G and Graves, Alex and Riedmiller, Martin and Fidjeland, Andreas K and Ostrovski, Georg and others},
journal={Nature},
volume={518},
number={7540},
pages={529--533},
year={2015},
publisher={Nature Research}
}
@article{lillicrap2015continuous,
title={Continuous control with deep reinforcement learning},
author={Lillicrap, Timothy P and Hunt, Jonathan J and Pritzel, Alexander and Heess, Nicolas and Erez, Tom and Tassa, Yuval and Silver, David and Wierstra, Daan},
journal={arXiv preprint arXiv:1509.02971},
year={2015}
}
flux/flux_forward_inverse_robot.jl
+ 18
− 5
View file @ e117c2da
# error("Kanske är forward/inverse regularization bättre på simulering än forward only. Inget signifikant syns vid prediktion.")
@everywhere begin @everywhere begin
# @everywhere include("/var/tmp/fredrikb/v0.6/DynamicMovementPrimitives/src/two_link.jl") # @everywhere include("/var/tmp/fredrikb/v0.6/DynamicMovementPrimitives/src/two_link.jl")
# @everywhere include(Pkg.dir("DynamicMovementPrimitives","src","two_link.jl")) # @everywhere include(Pkg.dir("DynamicMovementPrimitives","src","two_link.jl"))
...@@ -9,7 +10,7 @@ ...@@ -9,7 +10,7 @@
using Flux: back!, truncate!, treelike, train!, mse, testmode!, combine, params using Flux: back!, truncate!, treelike, train!, mse, testmode!, combine, params
using Flux.Optimise: weightdecay, Param, optimiser, RMSProp using Flux.Optimise: weightdecay, Param, optimiser, RMSProp
N = 2000 N = 500
n = 2 n = 2
ns = n ns = n
h = 0.02 h = 0.02
...@@ -36,8 +37,8 @@ ...@@ -36,8 +37,8 @@
sol = solve(prob,Tsit5(),reltol=1e-8,abstol=1e-8) sol = solve(prob,Tsit5(),reltol=1e-8,abstol=1e-8)
x = hcat(sol(t)...) x = hcat(sol(t)...)
# y = hcat([time_derivative(x[:,t], u[:,t])[3:4] for t in 1:N]...) # y = hcat([time_derivative(x[:,t], u[:,t])[3:4] for t in 1:N]...)
y = diff(x[1:2,:],2) y = x[3:4,2:N+1]#diff(x[1:2,:],2)
x = x[1:2,1:N] x = [x[1:2,1:N]; x[3:4,1:N]]
u = u[:,1:N] u = u[:,1:N]
# x[1:2,:] = x[1:2,:].% 2π # x[1:2,:] = x[1:2,:].% 2π
...@@ -52,6 +53,18 @@ ...@@ -52,6 +53,18 @@
x,u,y x,u,y
end end
function simulate(m,x)
Flux.testmode!.(m, true)
xsim = copy(x)
for t = 2:size(x,2)
xsimt = m(xsim[:,t-1])
xsim[3:4,t] = mean(xsimt).data
xsim[1:2,t] = xsim[1:2,t-1] + h*xsim[3:4,t-1]
end
Flux.testmode!.(m, false)
xsim
end
function fit_model(opt, loss, models, x, y, xv, yv; function fit_model(opt, loss, models, x, y, xv, yv;
iters = 2000, iters = 2000,
doplot = true, doplot = true,
...@@ -79,7 +92,7 @@ ...@@ -79,7 +92,7 @@
testmode!(ms) testmode!(ms)
l = loss(x, y).data[1] l = loss(x, y).data[1]
push!(trace,iter,l) push!(trace,iter,l)
if iter % 5 == 0 if iter % 10 == 0
push!(vtrace, iter, loss(xv,yv).data[1]) push!(vtrace, iter, loss(xv,yv).data[1])
push!(ftrace, iter, loss_forwardo(xv,yv).data[1]) push!(ftrace, iter, loss_forwardo(xv,yv).data[1])
push!(itrace, iter, loss_inverse(x,y).data[1]) push!(itrace, iter, loss_inverse(x,y).data[1])
...@@ -87,7 +100,7 @@ ...@@ -87,7 +100,7 @@
println("Iter: $iter, Loss: ", l) println("Iter: $iter, Loss: ", l)
plot(trace,reuse=true,show=false, lab="Train", layout=5, subplot=1, size=(1400,1000)) plot(trace,reuse=true,show=false, lab="Train", layout=5, subplot=1, size=(1400,1000))
plot!(vtrace,show=false, lab="Validation", subplot=1) plot!(vtrace,show=false, lab="Validation", subplot=1)
plot!(ftrace,show=false, lab="Forward validation", subplot=1) plot!(ftrace,show=false, lab="Forward validation", subplot=1, yscale=:log10)
# plot!(itrace,show=false, lab="Inverse validation", subplot=1) # plot!(itrace,show=false, lab="Inverse validation", subplot=1)
plot!(y', subplot=2:3, title="Forward model") plot!(y', subplot=2:3, title="Forward model")
plot!(my(x).data', l=:dash, subplot=2:3) plot!(my(x).data', l=:dash, subplot=2:3)
......
flux/nn_prior/nn_prior.jl
+ 8
− 8
View file @ e117c2da
...@@ -91,19 +91,19 @@ results1 = pmap(1:n_bootstrap) do it ...@@ -91,19 +91,19 @@ results1 = pmap(1:n_bootstrap) do it
println("Done: ", it) println("Done: ", it)
results results
end end
plotresults(results1)
Jm, Js, Jtrue = all_jacobians(results1) Jm, Js, Jtrue = all_jacobians(results1)
# Jtrue .-= [I zeros(4,2][:]' # Jtrue .-= [I zeros(4,2][:]'
l = @layout [ [a{1.0w}; b{1.0w}] c{0.5w}]
plot(plot_jacobians(Jm, Js, Jtrue), plot(plot_jacobians(Jm, Js, Jtrue),
plot_eigvals(results1, systype == DiffSystem()), plotresults(results1),
plotresults(results1)) plot_eigvals(results1, false),layout=l)
gui() gui()
plot(resue=false)
systype = DiffSystem() systype = DiffSystem()
seed = 1 seed = 1
x,y,u = generate_data(sys, systype, seed) x,y,u = generate_data(sys, systype, seed)
...@@ -121,13 +121,13 @@ results2 = pmap(1:n_bootstrap) do it ...@@ -121,13 +121,13 @@ results2 = pmap(1:n_bootstrap) do it
println("Done: ", it) println("Done: ", it)
results results
end end
plotresults(results2)
Jm, Js, Jtrue = all_jacobians(results2) Jm, Js, Jtrue = all_jacobians(results2)
# Jtrue .-= [I zeros(4,2][:]' # Jtrue .-= [I zeros(4,2][:]'
plot(plot_jacobians(Jm, Js, Jtrue, systype == DiffSystem()), l = @layout [ [a{1.0w}; b{1.0w}] c{0.5w}]
plot_eigvals(results2, systype == DiffSystem()), plot(plot_jacobians(Jm, Js, Jtrue, true),
plotresults(results2)) plotresults(results2),
plot_eigvals(results2, true), layout=l)
gui() gui()
......
flux/nn_prior/utilities.jl
+ 17
− 12
View file @ e117c2da
...@@ -97,19 +97,24 @@ function plot_jacobians(Jm, Js, Jtrue, shift=false) ...@@ -97,19 +97,24 @@ function plot_jacobians(Jm, Js, Jtrue, shift=false)
end end
function plot_eigvals(results, shift=false) function plot_eigvals(results, shift=false)
@unpack xv,uv,sys = results[1] @unpack x,u,sys = results[1]
N = size(xv,2) N = size(x,2)
plot() plot(layout=(2,1), ratio=:equal)
for evalpoint = 1:10:N for evalpoint = 1:10:N
e = eigvals(true_jacobian(sys,evalpoint,xv,uv)[1:4,1:4]) e = eigvals(true_jacobian(sys,evalpoint,x,u)[1:4,1:4])
scatter!(real.(e), imag.(e), c=:red, show=false) scatter!(real.(e), imag.(e), c=:red, show=false, subplot=1)
e = log.(e)/sys.h
e = eigvals(jacobian(models(results), xv[:,evalpoint])[1][1:4,1:4]) scatter!(real.(e), imag.(e), c=:red, show=false, subplot=2, legend=false)
scatter!(real.(e+shift), imag.(e), c=:blue, show=false)
e = eigvals(jacobian(models(results), x[:,evalpoint])[1][1:4,1:4])
e .+= shift
scatter!(real.(e), imag.(e), c=:blue, show=false, subplot=1)
e = log.(e)/sys.h
scatter!(real.(e), imag.(e), c=:blue, show=false, subplot=2, legend=false)
plot!(title="Eigenvalue spectrum") plot!(title="Eigenvalue spectrum")
end end
phi = linspace(0,2pi,300) phi = linspace(0,2pi,300)
plot!(real.(exp.(phi.*im)), imag.(exp.(phi.*im)), show=true, legend=false, c=:black, l=:dash) plot!(real.(exp.(phi.*im)), imag.(exp.(phi.*im)), show=true, legend=false, c=:black, l=:dash, subplot=1)
end end
...@@ -122,7 +127,7 @@ end ...@@ -122,7 +127,7 @@ end
function plotresults(results) function plotresults(results)
@unpack x,u,y,xv,uv,yv,systype = results[1] @unpack x,u,y,xv,uv,yv,systype = results[1]
fig = plot(xv[1:4,:]', lab="True", layout=4) fig = plot(x[1:4,:]', lab="True", layout=4)
ms = models(results) ms = models(results)
testmode!.(ms, true) testmode!.(ms, true)
plot_prediction(fig, results) plot_prediction(fig, results)
...@@ -135,7 +140,7 @@ end ...@@ -135,7 +140,7 @@ end
function plot_prediction(fig, results) function plot_prediction(fig, results)
@unpack x,u,y,xv,uv,yv,systype = results[1] @unpack x,u,y,xv,uv,yv,systype = results[1]
ms = models(results) ms = models(results)
yh, bounds = predict(systype, ms, xv) yh, bounds = predict(systype, ms, x)
for i = 1:4 for i = 1:4
plot!(fig, yh[i,:], fillrange = getindex.(bounds,i,:), fillalpha=0.3, subplot=i, lab="Prediction") plot!(fig, yh[i,:], fillrange = getindex.(bounds,i,:), fillalpha=0.3, subplot=i, lab="Prediction")
end end
...@@ -144,7 +149,7 @@ end ...@@ -144,7 +149,7 @@ end
function plot_simulation(fig, results) function plot_simulation(fig, results)
@unpack x,u,y,xv,uv,yv,systype = results[1] @unpack x,u,y,xv,uv,yv,systype = results[1]
ms = models(results) ms = models(results)
yh = simulate(systype, ms, xv) yh = simulate(systype, ms, x)
for i = 1:4 for i = 1:4
plot!(fig, yh[i,:], subplot=i, lab="Simulation") plot!(fig, yh[i,:], subplot=i, lab="Simulation")
end end
......
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Please register or to comment