Optimizing More than one Param

[ramtinkh]

I'm using NelderMead for my optimizing problem. I wanted to use the struct below.

impl CostFunction for MyStruct
{
    type Param = Vec<f64>;
    type Output = f64;

    fn cost(&self, input: &Self::Param) -> Result<Self::Output, Error> {
        let out = input[0].cos() + input[0].sin() * (input[1] / 100.0);
        Ok(out)
    }
}

Then I tried to define my solver like the code below. despite the documents of NelderMead for optimizing 2 parameters I should define 3 initial points.

let solver = NelderMead::new(vec![vec![0_f64, 70.0], vec![1_f64, 80.0], vec![2_f64, 90.0]]).with_sd_tolerance(vec![0.0001, 1.0]).unwrap();

after running this code I got errors mentioned here:

the trait `argmin_math::ArgminMul<_, Vec<f64>>` is not implemented for `Vec<f64>

What do you recommend me to do? Is it even possible for your CostFunction Trait to accept more than one parameter? Because vector doesn't implement ArgminMul, ...
Even when I copy the example in the document I got these errors.

use argmin::solver::neldermead::NelderMead;
    use argmin::core::Error;
    fn main() -> Result<(), Error> {
    let vec_of_parameters = vec![vec![1.0], vec![2.0], vec![3.0]];
    let nm: NelderMead<Vec<f64>, f64> =
        NelderMead::new(vec_of_parameters).with_sd_tolerance(1e-6)?;
    Ok(())
    }

[ramtinkh]

Even the example in this link failed with this error. Any ideas?

the trait `argmin_math::ArgminMul<_, ArrayBase<OwnedRepr<{float}>, Dim<[usize; 1]>>>` is not implemented for `ArrayBase<OwnedRepr<{float}>, Dim<[usize; 1]>>`

[GermanHeim]

Hello,

In your code you have .with_sd_tolerance(vec![0.0001, 1.0]) but sd_tolerance should be a float (https://docs.rs/argmin/latest/argmin/solver/neldermead/struct.NelderMead.html)

The example works for me:

OptimizationResult:
    Solver:        Nelder-Mead method
    param (best):  [1.0033431432617184, 1.0073634384738885], shape=[2], strides=[1], layout=CFcf (0xf), const ndim=1 
    cost (best):   0.00005552892269447157
    iters (best):  50
    iters (total): 54
    termination:   Solver converged
    time:          1.8594ms

Modifying the code inside the example to take your function also worked for me:

OptimizationResult:
    Solver:        Nelder-Mead method
    param (best):  [4.023193359375, 110.23193359375]
    cost (best):   -1.486640128657184
    iters (best):  11
    iters (total): 15
    termination:   Solver converged
    time:          677.1µs

Code:

use argmin::{
    core::{observers::ObserverMode, CostFunction, Error, Executor},
    solver::neldermead::NelderMead,
};
use argmin_observer_slog::SlogLogger;

struct MyStruct {}

impl CostFunction for MyStruct {
    type Param = Vec<f64>;
    type Output = f64;

    fn cost(&self, input: &Self::Param) -> Result<Self::Output, Error> {
        let out = input[0].cos() + input[0].sin() * (input[1] / 100.0);
        Ok(out)
    }
}

fn run() -> Result<(), Error> {
    // Define cost function
    let cost = MyStruct {};

    let solver = NelderMead::new(vec![
        vec![0_f64, 70.0],
        vec![1_f64, 80.0],
        vec![2_f64, 90.0],
    ])
    .with_sd_tolerance(0.0001)?;

    // Run solver
    let res = Executor::new(cost, solver)
        .configure(|state| state.max_iters(100))
        .add_observer(SlogLogger::term(), ObserverMode::Always)
        .run()?;

    // Print result
    println!("{res}");
    Ok(())
}

fn main() {
    if let Err(ref e) = run() {
        println!("{e}");
        std::process::exit(1);
    }
}

In my cargo.toml I have the following dependencies:

[dependencies]
argmin = { version = "0.10.0" }
argmin-math = { version = "0.4.0", features = ["ndarray_latest-nolinalg"] }
argmin-observer-slog = { version = "*" }
argmin_testfunctions = { version = "*" }
ndarray = "0.15.6" # Only useful for the example, not your problem

Could you check if that works for you?

[ramtinkh]

When I updated the code and run, it still return the following errors.

the trait `argmin_math::ArgminMul<_, Vec<f64>>` is not implemented for `Vec<f64>

Then I updated the dependencies from the details mentioned below to your dependencies.

[dependencies]
argmin = "0.10.0"
argmin-observer-slog = "0.1.0"
argmin_testfunctions = "0.2.0"
num-traits = "0.2.19"
ndarray = "0.15.6"

And now it works. Thank you so much.
But I didn't understand which dependency creates the previous error. If you had any idea let me know.
Thanks for your time.

[GermanHeim]

What were your previous dependencies?

[ramtinkh]

I’ve added my previous dependencies in the previous comment.

[GermanHeim]

Oh, sorry, I didn't read that right.
I think you were missing argmin-math.