# Experiment Design

Originally Contributed by: Arpit Bhatia, Chris Coey

This tutorial covers experiment design examples (D-optimal, A-optimal, and E-optimal) from section 7.5 of the book Convex Optimization by Boyd and Vandenberghe.

The tutorial uses the following packages

using JuMP
import SCS
import LinearAlgebra
import Random

We set a seed so the random numbers are repeatable:

Random.seed!(1234)
Random.MersenneTwister(UInt32[0x000004d2], Random.DSFMT.DSFMT_state(Int32[-1393240018, 1073611148, 45497681, 1072875908, 436273599, 1073674613, -2043716458, 1073445557, -254908435, 1072827086  …  -599655111, 1073144102, 367655457, 1072985259, -1278750689, 1018350124, -597141475, 249849711, 382, 0]), [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0  …  0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], UInt128[0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000  …  0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000, 0x00000000000000000000000000000000], 1002, 0)

## Relaxed Experiment Design Problem

The basic experiment design problem is as follows.

Given the menu of possible choices for experiments, $v_{1}, \ldots, v_{p}$, and the total number $m$ of experiments to be carried out, choose the numbers of each type of experiment, i.e., $m_{1}, \ldots, m_{p}$ to make the error covariance $E$ small (in some sense).

The variables $m_{1}, \ldots, m_{p}$ must, of course, be integers and sum to $m$ the given total number of experiments. This leads to the optimization problem:

\begin{aligned} \min\left(\mathrm{w.r.t.} \mathbf{S}_{+}^{n}\right) & E=\left(\sum_{j=1}^{p} m_{j} v_{j} v_{j}^{T}\right)^{-1} \\ \text{subject to} & m_{i} \geq 0 \\ & \sum\limits_{i=1}^p m_{i} = m \\ & m_{i} \in \mathbb{Z},\quad i=1,\ldots,p \end{aligned}

The basic experiment design problem can be a hard combinatorial problem when $m$, the total number of experiments, is comparable to $n$, since in this case the $m_{i}$ are all small integers.

In the case when $m$ is large compared to $n$, however, a good approximate solution can be found by ignoring, or relaxing, the constraint that the $m_{i}$ are integers.

Let $\lambda_{i}=m_{i} / m,$ which is the fraction of the total number of experiments for which $a_{j}=v_{i},$ or the relative frequency of experiment $i$. We can express the error covariance in terms of $\lambda_{i}$ as:

$$$E=\frac{1}{m}\left(\sum_{i=1}^{p} \lambda_{i} v_{i} v_{i}^{T}\right)^{-1}$$$

The vector $\lambda \in \mathbf{R}^{p}$ satisfies $\lambda \succeq 0, \mathbf{1}^{T} \lambda=1,$ and also, each $\lambda_{i}$ is an integer multiple of $1 / m$. By ignoring this last constraint, we arrive at the problem:

\begin{aligned} \min\left(\mathrm{w.r.t.} \mathbf{S}_{+}^{n}\right) & E=(1 / m)\left(\sum_{i=1}^{p} \lambda_{i} v_{i} v_{i}^{T}\right)^{-1} \\ \text{subject to:}\quad & \lambda \succeq 0 \\ & \mathbf{1}^{T} \lambda=1 \end{aligned}

Several scalarizations have been proposed for the experiment design problem, which is a vector optimization problem over the positive semidefinite cone.

q = 4 # dimension of estimate space
p = 8 # number of experimental vectors
nmax = 3 # upper bound on lambda
n = 12

V = randn(q, p)

eye = Matrix{Float64}(LinearAlgebra.I, q, q);

## A-optimal design

In A-optimal experiment design, we minimize tr $E$, the trace of the covariance matrix. This objective is simply the mean of the norm of the error squared:

$$$\mathbf{E}\|e\|_{2}^{2}=\mathbf{E} \operatorname{tr}\left(e e^{T}\right)=\operatorname{tr} E$$$

The A-optimal experiment design problem in SDP form is

\begin{aligned} \min & \mathbf{1}^{T} u \\ \text{subject to} & \left[\begin{aligned}{\sum_{i=1}^{p} \lambda_{i} v_{i} v_{i}^{T}} & {e_{k}} \\ {e_{k}^{T}} & {u_{k}}\end{aligned}\right] \succeq 0, \quad k=1, \ldots, n \\ & \lambda \succeq 0 \\ & \mathbf{1}^{T} \lambda=1 \end{aligned}
aOpt = Model(SCS.Optimizer)
set_silent(aOpt)
@variable(aOpt, np[1:p], lower_bound = 0, upper_bound = nmax)
@variable(aOpt, u[1:q], lower_bound = 0)
@constraint(aOpt, sum(np) <= n)
for i = 1:q
matrix = [
V * LinearAlgebra.diagm(0 => np ./ n) * V' eye[:, i];
eye[i, :]' u[i]
]
@SDconstraint(aOpt, matrix >= 0)
end
@objective(aOpt, Min, sum(u))
optimize!(aOpt)
objective_value(aOpt)
5.041247589265131
value.(np)
8-element Array{Float64,1}:
1.7479360436953184
1.115313540064675
1.8899140021302877e-6
1.6619566418994247
2.999996940001866
0.8414161117660524
1.3825673960019045
2.2508040064854136

## E-optimal design

In $E$ -optimal design, we minimize the norm of the error covariance matrix, i.e. the maximum eigenvalue of $E$.

Since the diameter (twice the longest semi-axis) of the confidence ellipsoid $\mathcal{E}$ is proportional to $\|E\|_{2}^{1 / 2}$, minimizing $\|E\|_{2}$ can be interpreted geometrically as minimizing the diameter of the confidence ellipsoid.

E-optimal design can also be interpreted as minimizing the maximum variance of $q^{T} e$, over all $q$ with $\|q\|_{2}=1$. The E-optimal experiment design problem in SDP form is:

\begin{aligned} \min & t \\ \text{subject to} & \sum_{i=1}^{p} \lambda_{i} v_{i} v_{i}^{T} \succeq t I \\ & \lambda \succeq 0 \\ & \mathbf{1}^{T} \lambda=1 \end{aligned}
eOpt = Model(SCS.Optimizer)
set_silent(eOpt)
@variable(eOpt, 0 <= np[1:p] <= nmax)
@variable(eOpt, t)
@SDconstraint(eOpt, V * LinearAlgebra.diagm(0 => np ./ n) * V' - (t .* eye) >= 0)
@constraint(eOpt, sum(np) <= n)
@objective(eOpt, Max, t)
optimize!(eOpt)
objective_value(eOpt)
0.4489430761395082
value.(np)
8-element Array{Float64,1}:
3.0000033587011705
0.6752098189054311
-2.1880963065487475e-6
1.0458573703123515
2.9999992832228677
1.7869721644137855
0.30150709168864204
2.1904604684506817

## D-optimal design

The most widely used scalarization is called $D$ -optimal design, in which we minimize the determinant of the error covariance matrix $E$. This corresponds to designing the experiment to minimize the volume of the resulting confidence ellipsoid (for a fixed confidence level). Ignoring the constant factor $1 / m$ in $E$, and taking the logarithm of the objective, we can pose this problem as convex optimization problem:

\begin{aligned} \min & \log \operatorname{det}\left(\sum_{i=1}^{p} \lambda_{i} v_{i} v_{i}^{T}\right)^{-1} \\ \text{subject to} & \lambda \succeq 0 \\ & \mathbf{1}^{T} \lambda=1 \end{aligned}
dOpt = Model(SCS.Optimizer)
set_silent(dOpt)
@variable(dOpt, np[1:p], lower_bound = 0, upper_bound = nmax)
@variable(dOpt, t)
@objective(dOpt, Max, t)
@constraint(dOpt, sum(np) <= n)
E = V * LinearAlgebra.diagm(0 => np ./ n) * V'
@constraint(
dOpt,
[t, 1, (E[i, j] for i in 1:q for j in 1:i)...] in MOI.LogDetConeTriangle(q)
)
optimize!(dOpt)
objective_value(dOpt)
0.19014655971032338
value.(np)
8-element Array{Float64,1}:
-1.4105081413253679e-6
2.5673993850093133
-8.699434116245897e-7
0.2627510335334194
2.9434767950403455
2.3925302087681466
2.8369741480581023
0.9968841960241067