Meet the Brave New Algorithm

Literature-inspired metaphor

Stratified population evolutionary algorithm

α caste → with itself, crossover + mutation

β only with α → crossover + mutation

γ, δ, ε → only mutation; γ → hillclimbing

Open source, written in Julia!

Very fine control of the exploration/exploitation tradeoff

Get it from CeciMerelo/BraveNewAlgorithm.jl at GitHub

It's only possible to measure energy spent by the system while our program is running

E(Workload) = E(measured) - E(baseline)

Baseline includes operating context + runtime overhead

But we are interested in our algorithm only

E(Workload) = E(measured) - E(baseline)

So, again, can we measure energy consumption in population-based metaheuristics?

Maybe not? Check out these baseline measures

Baseline is a fleeting thing...

We need to avoid variability

Or rather to track it

Sequential mixed strategy → 1 Baseline, 1 Workload

Again choppy waters... But at least the ships keep together

Still not ideal for measuring small or medium differences

Algorithm configuration would be out

Let's go for a sandwich

Baseline sandwich ⇒ a workload between 2 baseline runs

E(Workload) = E(measured) - (E(baseline-before)-E(baseline-after))/2

Also, we make 5 blocks of experiments

Measuring over diverse operating contexts

Even so, this is what we have

📊 Can we spot differences depending on problem dimension/population sizes?