Rebuild DiffEqUncertainty/01-expectation_introduction.jmd, DiffEqUncertainty/02-AD_and_optimization.jmd

markdown/DiffEqUncertainty/01-expectation_introduction.md

@@ -2,15 +2,6 @@
 author: "Adam Gerlach"
 title: "An Intro to Expectations via DiffEqUncertainty.jl"
 ---
-<!-- 
-using Pkg, SciMLTutorials
-cd(joinpath(dirname(pathof(SciMLTutorials)), ".."))
-Pkg.pkg"activate ."
-Pkg.pkg"instantiate"
-SciMLTutorials.weave_file("DiffEqUncertainty","01-expectation_introduction.jmd",(:github,))
-SciMLTutorials.weave_file("DiffEqUncertainty","01-expectation_introduction.jmd") 
--->
-
 
 
 ## System Model
@@ -120,7 +111,7 @@ mean([g(sol) for sol in ensemblesol])
 
 ````
 1-element Array{Float64,1}:
- -0.003406754846664461
+ 0.005073268080398201
 ````
 
 
@@ -137,7 +128,7 @@ expectation(g, prob, u0_dist, p, MonteCarlo(), Tsit5(); trajectories=100000)
 
 ````
 1-element Array{Float64,1}:
- 0.007225918038175185
+ 0.005772652448278111
 ````
 
 
@@ -191,9 +182,9 @@ We see that for this case the `Koopman()` algorithm produces a more accurate sol
 
 
 ````
-2.078746 seconds (79.60 M allocations: 7.168 GiB, 71.29% gc time)
+2.215703 seconds (79.60 M allocations: 7.168 GiB, 73.13% gc time)
 1-element Array{Float64,1}:
- 0.0002649264693847175
+ 0.00412338430030202
 ````
 
 
@@ -204,7 +195,7 @@ We see that for this case the `Koopman()` algorithm produces a more accurate sol
 
 
 ````
-0.000824 seconds (12.28 k allocations: 1.112 MiB)
+0.000801 seconds (12.28 k allocations: 1.112 MiB)
 u: 1-element Array{Float64,1}:
  0.0
 ````
@@ -222,9 +213,9 @@ u0_dist = [Uniform(0.0,10.0)]
 
 
 ````
-1.082617 seconds (79.59 M allocations: 7.168 GiB, 45.45% gc time)
+0.653305 seconds (79.60 M allocations: 7.168 GiB)
 1-element Array{Float64,1}:
- 1.5058180695041825
+ 1.504424534047735
 ````
 
 
@@ -237,7 +228,7 @@ and
 
 
 ````
-0.004236 seconds (14.04 k allocations: 1.221 MiB)
+0.004519 seconds (14.04 k allocations: 1.221 MiB)
 1.5059722133001539
 ````
 
@@ -596,9 +587,8 @@ using Quadrature, Cuba
 
 
 ````
-Error: ArgumentError: Package Quadrature not found in current path:
-- Run `import Pkg; Pkg.add("Quadrature")` to install the Quadrature package
-.
+Error: ArgumentError: Package Cuba not found in current path:
+- Run `import Pkg; Pkg.add("Cuba")` to install the Cuba package.
 ````
 
 
@@ -847,74 +837,9 @@ Status `/builds/JuliaGPU/DiffEqTutorials.jl/tutorials/DiffEqUncertainty/Project.
 [ef61062a-5684-51dc-bb67-a0fcdec5c97d] DiffEqUncertainty 1.5.0
 [0c46a032-eb83-5123-abaf-570d42b7fbaa] DifferentialEquations 6.15.0
 [31c24e10-a181-5473-b8eb-7969acd0382f] Distributions 0.23.8
+[f6369f11-7733-5829-9624-2563aa707210] ForwardDiff 0.10.12
 [76087f3c-5699-56af-9a33-bf431cd00edd] NLopt 0.6.0
 [1dea7af3-3e70-54e6-95c3-0bf5283fa5ed] OrdinaryDiffEq 5.42.1
 [91a5bcdd-55d7-5caf-9e0b-520d859cae80] Plots 1.5.8
+[67601950-bd08-11e9-3c89-fd23fb4432d2] Quadrature 1.3.0
 ```
-
-
-
-
-<!--
-## Batch + nout
-````julia
-using DiffEqGPU
-
-function f(du, u,p,t) 
-    @inbounds begin
-        du[1] = p[1]*u[1];
-    end
-    nothing
-end
-
-u0 = Float32[10.0]
-p = Float32[-0.3]
-tspan = (0.0f0,10.0f0)
-prob = ODEProblem(f,u0,tspan,p)
-
-g(sol) = [sol(4.0)[1], sol(6.0)[1]]
-
-u0_dist = [truncated(Normal(3.0f0,2.0f0),-5f0,11f0)]
-p_dist = [truncated(Normal(-.7f0, .1f0), -1f0,0f0)]
-
-expectation(g, prob, u0_dist, p_dist, Koopman(), Tsit5(), EnsembleGPUArray(); 
-            quadalg = CubaSUAVE(), batch=1000, nout=2)[1]
-````
-
-
-````
-Error: MethodError: no method matching __solvebp_call(::DiffEqBase.Quadratu
-reProblem{true,Array{Float32,1},DiffEqUncertainty.var"#13#26"{DiffEqUncerta
-inty.var"#20#33",DiffEqUncertainty.var"#21#34",Base.Iterators.Pairs{Union{}
-,Union{},Tuple{},NamedTuple{(),Tuple{}}},typeof(Main.##WeaveSandBox#496.g),
-DiffEqBase.ODEProblem{Array{Float32,1},Tuple{Float32,Float32},true,Array{Fl
-oat32,1},DiffEqBase.ODEFunction{true,typeof(Main.##WeaveSandBox#496.f),Line
-arAlgebra.UniformScaling{Bool},Nothing,Nothing,Nothing,Nothing,Nothing,Noth
-ing,Nothing,Nothing,Nothing,Nothing,Nothing,Nothing},Base.Iterators.Pairs{U
-nion{},Union{},Tuple{},NamedTuple{(),Tuple{}}},DiffEqBase.StandardODEProble
-m},Tuple{OrdinaryDiffEq.Tsit5,DiffEqGPU.EnsembleGPUArray},Int64,Array{Bool,
-1},Array{Distributions.Truncated{Distributions.Normal{Float32},Distribution
-s.Continuous,Float32},1},Array{Float32,1}},Array{Float32,1},Array{Float32,1
-},Base.Iterators.Pairs{Union{},Union{},Tuple{},NamedTuple{(),Tuple{}}}}, ::
-Quadrature.CubaSUAVE, ::Quadrature.ReCallVJP{Quadrature.ZygoteVJP}, ::Array
-{Float32,1}, ::Array{Float32,1}, ::Array{Float32,1}; reltol=0.01, abstol=0.
-01, maxiters=1000000)
-Closest candidates are:
-  __solvebp_call(::DiffEqBase.QuadratureProblem, !Matched::Quadrature.QuadG
-KJL, ::Any, ::Any, ::Any, ::Any, !Matched::Any...; reltol, abstol, maxiters
-, kwargs...) at /builds/JuliaGPU/DiffEqTutorials.jl/.julia/packages/Quadrat
-ure/L8aMP/src/Quadrature.jl:70
-  __solvebp_call(::DiffEqBase.QuadratureProblem, !Matched::Quadrature.HCuba
-tureJL, ::Any, ::Any, ::Any, ::Any, !Matched::Any...; reltol, abstol, maxit
-ers, kwargs...) at /builds/JuliaGPU/DiffEqTutorials.jl/.julia/packages/Quad
-rature/L8aMP/src/Quadrature.jl:87
-  __solvebp_call(::DiffEqBase.QuadratureProblem, !Matched::Quadrature.VEGAS
-, ::Any, ::Any, ::Any, ::Any, !Matched::Any...; reltol, abstol, maxiters, k
-wargs...) at /builds/JuliaGPU/DiffEqTutorials.jl/.julia/packages/Quadrature
-/L8aMP/src/Quadrature.jl:113
-````
-
-
-
-
--->

markdown/DiffEqUncertainty/02-AD_and_optimization.md

@@ -2,16 +2,6 @@
 author: "Adam Gerlach"
 title: "Optimization Under Uncertainty with DiffEqUncertainty.jl"
 ---
-<!-- 
-using Pkg, SciMLTutorials
-cd(joinpath(dirname(pathof(SciMLTutorials)), ".."))
-Pkg.pkg"activate ."
-Pkg.pkg"instantiate"
-SciMLTutorials.weave_file("DiffEqUncertainty","02-AD_and_optimization.jmd",(:github,))
-SciMLTutorials.weave_file("DiffEqUncertainty","02-AD_and_optimization.jmd") 
-
--->
-
 
 
 This tutorial gives and overview of how to leverage the efficient Koopman expectation method from DiffEqUncertainty to perform optimization under uncertainty. We demonstrate this by using a bouncing ball model with an uncertain model parameter. We also demonstrate its application to problems with probabilistic constraints, in particular a special class of constraints called chance constraints. 
@@ -138,7 +128,7 @@ mean_ensemble = mean([obs(sol) for sol in ensemblesol])
 
 
 ````
-35.986155065108754
+35.70585036601521
 ````
 
 
@@ -171,18 +161,6 @@ First, we load the required packages and define our loss function
 
 ````julia
 using NLopt, DiffEqSensitivity, ForwardDiff
-````
-
-
-````
-Error: ArgumentError: Package ForwardDiff not found in current path:
-- Run `import Pkg; Pkg.add("ForwardDiff")` to install the ForwardDiff packa
-ge.
-````
-
-
-
-````julia
 
 make_u0(θ) = [θ[1],θ[2],θ[3], 0.0]
 
@@ -233,7 +211,8 @@ opt.min_objective = 𝔼_loss_nlopt
 
 
 ````
-(0.0, [-1.0, 2.0, 50.0], :FORCED_STOP)
+(0.0838151729557601, [-0.05713109086715892, 2.4366742783682063, 49.99835957
+072762], :XTOL_REACHED)
 ````
 
 
@@ -274,8 +253,9 @@ using BenchmarkTools
 
 
 ````
-119.826 μs (8 allocations: 448 bytes)
-(0.0, [-1.0, 2.0, 50.0], :FORCED_STOP)
+129.798 ms (851126 allocations: 63.07 MiB)
+(0.0838151729557601, [-0.05713109086715892, 2.4366742783682063, 49.99835957
+072762], :XTOL_REACHED)
 ````
 
 
@@ -363,7 +343,7 @@ expectation(constraint_obs, prob, make_u0(minx), p_uncertain, Koopman(), Tsit5()
 
 
 ````
-0.0028823461332289553
+0.9910305871387488
 ````
 
 
@@ -410,7 +390,8 @@ inequality_constraint!(opt,𝔼_constraint_nlopt, 1e-5)
 
 
 ````
-(0.0, [-1.0, 2.0, 50.0], :FORCED_STOP)
+(40.28996781875263, [-0.8176321801341233, 2.000952669391841, 50.0], :XTOL_R
+EACHED)
 ````
 
 
@@ -425,20 +406,20 @@ The probability of impacting the wall is now
 
 
 ````
-0.0028823461332289553
+0.010004498174330818
 ````
 
 
 
 
 We can check if this is within tolerance by
 ````julia
-isapprox(λ, 0.01, atol=1e-5)
+λ - 0.01 <= 1e-5
 ````
 
 
 ````
-false
+true
 ````
 
 
@@ -471,13 +452,6 @@ end
 ![](figures/02-AD_and_optimization_21_1.png)
 
 
-
-
-
-## Conclusion
-
-
-
 ## Appendix
 
  This tutorial is part of the SciMLTutorials.jl repository, found at: <https://github.com/SciML/SciMLTutorials.jl>.
@@ -517,7 +491,9 @@ Status `/builds/JuliaGPU/DiffEqTutorials.jl/tutorials/DiffEqUncertainty/Project.
 [ef61062a-5684-51dc-bb67-a0fcdec5c97d] DiffEqUncertainty 1.5.0
 [0c46a032-eb83-5123-abaf-570d42b7fbaa] DifferentialEquations 6.15.0
 [31c24e10-a181-5473-b8eb-7969acd0382f] Distributions 0.23.8
+[f6369f11-7733-5829-9624-2563aa707210] ForwardDiff 0.10.12
 [76087f3c-5699-56af-9a33-bf431cd00edd] NLopt 0.6.0
 [1dea7af3-3e70-54e6-95c3-0bf5283fa5ed] OrdinaryDiffEq 5.42.1
 [91a5bcdd-55d7-5caf-9e0b-520d859cae80] Plots 1.5.8
+[67601950-bd08-11e9-3c89-fd23fb4432d2] Quadrature 1.3.0
 ```

notebook/DiffEqUncertainty/01-expectation_introduction.ipynb

@@ -3,7 +3,7 @@
     {
       "cell_type": "markdown",
       "source": [
-        "<!-- \nusing Pkg, SciMLTutorials\ncd(joinpath(dirname(pathof(SciMLTutorials)), \"..\"))\nPkg.pkg\"activate .\"\nPkg.pkg\"instantiate\"\nSciMLTutorials.weave_file(\"DiffEqUncertainty\",\"01-expectation_introduction.jmd\",(:github,))\nSciMLTutorials.weave_file(\"DiffEqUncertainty\",\"01-expectation_introduction.jmd\") \n-->\n\n\n\n## System Model\n\nFirst, lets consider the following linear model.\n\n$$u' = p u$$"
+        "## System Model\n\nFirst, lets consider the following linear model.\n\n$$u' = p u$$"
       ],
       "metadata": {}
     },
@@ -552,29 +552,6 @@
         "The performance gains realized by leveraging batch GPU processing is problem dependent. In this case, the number of batch evaluations required to overcome the overhead of using the GPU exceeds the number of simulations required to converge to the quadrature solution."
       ],
       "metadata": {}
-    },
-    {
-      "cell_type": "markdown",
-      "source": [
-        "<!--\n## Batch + nout"
-      ],
-      "metadata": {}
-    },
-    {
-      "outputs": [],
-      "cell_type": "code",
-      "source": [
-        "using DiffEqGPU\n\nfunction f(du, u,p,t) \n    @inbounds begin\n        du[1] = p[1]*u[1];\n    end\n    nothing\nend\n\nu0 = Float32[10.0]\np = Float32[-0.3]\ntspan = (0.0f0,10.0f0)\nprob = ODEProblem(f,u0,tspan,p)\n\ng(sol) = [sol(4.0)[1], sol(6.0)[1]]\n\nu0_dist = [truncated(Normal(3.0f0,2.0f0),-5f0,11f0)]\np_dist = [truncated(Normal(-.7f0, .1f0), -1f0,0f0)]\n\nexpectation(g, prob, u0_dist, p_dist, Koopman(), Tsit5(), EnsembleGPUArray(); \n            quadalg = CubaSUAVE(), batch=1000, nout=2)[1]"
-      ],
-      "metadata": {},
-      "execution_count": null
-    },
-    {
-      "cell_type": "markdown",
-      "source": [
-        "-->"
-      ],
-      "metadata": {}
     }
   ],
   "nbformat_minor": 2,

notebook/DiffEqUncertainty/02-AD_and_optimization.ipynb

@@ -3,7 +3,7 @@
     {
       "cell_type": "markdown",
       "source": [
-        "<!-- \nusing Pkg, SciMLTutorials\ncd(joinpath(dirname(pathof(SciMLTutorials)), \"..\"))\nPkg.pkg\"activate .\"\nPkg.pkg\"instantiate\"\nSciMLTutorials.weave_file(\"DiffEqUncertainty\",\"02-AD_and_optimization.jmd\",(:github,))\nSciMLTutorials.weave_file(\"DiffEqUncertainty\",\"02-AD_and_optimization.jmd\") \n\n-->\n\n\n\nThis tutorial gives and overview of how to leverage the efficient Koopman expectation method from DiffEqUncertainty to perform optimization under uncertainty. We demonstrate this by using a bouncing ball model with an uncertain model parameter. We also demonstrate its application to problems with probabilistic constraints, in particular a special class of constraints called chance constraints. \n\n## System Model\nFirst lets consider a 2D bouncing ball, where the states are the horizontal position $x$, horizontal velocity $\\dot{x}$, vertical position $y$, and vertical velocity $\\dot{y}$. This model has two system parameters, acceleration due to gravity and coefficient of restitution (models energy loss when the ball impacts the ground). We can simulate such a system using `ContinuousCallback` as"
+        "This tutorial gives and overview of how to leverage the efficient Koopman expectation method from DiffEqUncertainty to perform optimization under uncertainty. We demonstrate this by using a bouncing ball model with an uncertain model parameter. We also demonstrate its application to problems with probabilistic constraints, in particular a special class of constraints called chance constraints. \n\n## System Model\nFirst lets consider a 2D bouncing ball, where the states are the horizontal position $x$, horizontal velocity $\\dot{x}$, vertical position $y$, and vertical velocity $\\dot{y}$. This model has two system parameters, acceleration due to gravity and coefficient of restitution (models energy loss when the ball impacts the ground). We can simulate such a system using `ContinuousCallback` as"
       ],
       "metadata": {}
     },
@@ -315,7 +315,7 @@
       "outputs": [],
       "cell_type": "code",
       "source": [
-        "isapprox(λ, 0.01, atol=1e-5)"
+        "λ - 0.01 <= 1e-5"
       ],
       "metadata": {},
       "execution_count": null
@@ -335,13 +335,6 @@
       ],
       "metadata": {},
       "execution_count": null
-    },
-    {
-      "cell_type": "markdown",
-      "source": [
-        "## Conclusion"
-      ],
-      "metadata": {}
     }
   ],
   "nbformat_minor": 2,

script/DiffEqUncertainty/01-expectation_introduction.jl

@@ -201,26 +201,3 @@ p_dist = [truncated(Normal(-.7f0, .1f0), -1f0,0f0)]
 using SciMLTutorials
 SciMLTutorials.tutorial_footer(WEAVE_ARGS[:folder],WEAVE_ARGS[:file])
 
-
-using DiffEqGPU
-
-function f(du, u,p,t) 
-    @inbounds begin
-        du[1] = p[1]*u[1];
-    end
-    nothing
-end
-
-u0 = Float32[10.0]
-p = Float32[-0.3]
-tspan = (0.0f0,10.0f0)
-prob = ODEProblem(f,u0,tspan,p)
-
-g(sol) = [sol(4.0)[1], sol(6.0)[1]]
-
-u0_dist = [truncated(Normal(3.0f0,2.0f0),-5f0,11f0)]
-p_dist = [truncated(Normal(-.7f0, .1f0), -1f0,0f0)]
-
-expectation(g, prob, u0_dist, p_dist, Koopman(), Tsit5(), EnsembleGPUArray(); 
-            quadalg = CubaSUAVE(), batch=1000, nout=2)[1]
-