Benchmarks

Create the most commonly used Dynamical Systems Benchmark systems

Getting started

Clone the directory, go to directory, open the package manager in the Julia REPL using "]", and add the package:

(@v1.8) pkg> add .

Or enter the package manager in the Juila REPL, and add the package with:

(@v1.8) pkg> add github_link

Generate Data

from directory do

julia --project main.jl

with arguments specified in settings/benchmark_defaults.json, i.e. you need a folder settings with benchmark_defaults.json in the parent directory

or in commandline with

julia --project main.jl --kwarg values

choose the benchmark system with --name

Usage in scripts

using Benchmarks

generate_benchmarks(parse_commandline())

that's it, adjust settings in settings/benchmark_defaults.json

If you want to create your own system follow the steps in template.jl

Little Warning, process noise increases computation time drastically

Implemented Systems

Standard Benchmarks

• standard_bursting: Bursting Neuron as used in all Durstewitz Lab papers
• standard_lorenz: Lorenz, with ρ = 28
• lorenz_limit_cycle: A limit cycle of the Lorenz System with ρ=24
• bursting_limit_cycle: The Bursting neuron limit cycle at gₙₘ₀ₐ=10

Regimes

• bursting_neuron_regimes: Several regimes of the bursting neuron

Non-Stationary Benchmarks

Lorenz

• ShrinkingLorenz : Starts with the chaotic attractor and shrinks plus shifts it a bit; ρ=28->23, σ=10->5, β=8/3->0.5 (transient Time should not be bigger than 50, as it leads to weird regimes)

• PaperLorenzBigChange : The ns system used in Patel et al. 2022 with a quick parameter change

• PaperLorenzSmallChange : The ns system used in Patel et al. 2022 with a slow parameter change

• ExplodingLorenz : Starts with a limit cycle and ends in the well known chaotic attractor; ρ=22->28 (not a very stable system, i.e. sensitive to initial conditions, makes it harder to train on)

• ShiftingLorenz : Starts with the chaotic attractor and shifts it "forward"; ρ=28->22 (limit cycle does not appear in time series, -> not great for snapshot comparison)

Bursting Neuron

The timescale is much longer than for the Lorenz, and thus needs a lot more time points

• RampUpBN : Starts with the a cycle, and adds bursting loops; g=2->4

• StopBurstBN : Starts with the complicated cycle at g=9.25 and ends in a limit cycle at g=10.15

Trial Benchmark Systems

• trial_lorenz : a lorenz with parameter shifting from 22->28 across trials
• split_lorenz : a specified ns lorenz split into num_trials

see docs for more information on what kwargs are availabe