1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156

use super::{LogLikelihood, Weighted};
use crate::error::Error;
use crate::error::Error::*;
use argmin::prelude::*;
use argmin::solver::neldermead::NelderMead;
use ndarray::RawData;
use num_traits::{Float, FloatConst, FromPrimitive};
use serde::{Deserialize, Serialize};
use std::convert::TryFrom;
use std::fmt::{Debug, Display};
use std::marker::PhantomData;

const NON_ZERO_DELTA: f64 = 0.05;
const ZERO_DELTA: f64 = 0.00025;

pub trait InitialSolvePoint<T> {
    fn initial_solve_point(&self) -> T;
}

impl<T, Distribution, W, D> InitialSolvePoint<Distribution> for Weighted<T, W, D>
where
    W: RawData,
    T: InitialSolvePoint<Distribution>,
{
    fn initial_solve_point(&self) -> Distribution {
        self.time.initial_solve_point()
    }
}

pub trait InitialNelderMeanSimplex<T> {
    fn initial_simplex(&self) -> Result<Vec<T>, Error>;
}

impl<F> InitialNelderMeanSimplex<Vec<F>> for [F]
where
    F: Float + FromPrimitive,
{
    fn initial_simplex(&self) -> Result<Vec<Vec<F>>, Error> {
        let initial_point: Vec<F> = self.into();

        let d = initial_point.len();
        let mut simplex = vec![initial_point; d + 1];
        for (index_within_point, point) in simplex.iter_mut().skip(1).enumerate() {
            if point[index_within_point] != F::zero() {
                let delta_multiple = 1.0 + NON_ZERO_DELTA;

                point[index_within_point] = F::from(delta_multiple)
                    .ok_or(NumericalConversion(delta_multiple))?
                    * point[index_within_point]
            } else {
                point[index_within_point] =
                    F::from(ZERO_DELTA).ok_or(NumericalConversion(ZERO_DELTA))?
            }
        }

        Ok(simplex)
    }
}

pub struct BaseFitter<S, D, F> {
    input_state: S,
    distribution: PhantomData<D>,
    float: PhantomData<F>,
    pub max_iterations: u64,
}

impl<S, D, F> BaseFitter<S, D, F> {
    pub fn new(data: S) -> Self {
        BaseFitter {
            input_state: data,
            distribution: PhantomData,
            float: PhantomData,
            max_iterations: 100,
        }
    }
}

impl<'f, S, D, F> ArgminOp for &'f BaseFitter<S, D, F>
where
    S: LogLikelihood<D, F>,
    D: for<'a> TryFrom<&'a [F], Error = Error>,
    F: Float + FloatConst + FromPrimitive + Debug + Display + Serialize + for<'de> Deserialize<'de>,
{
    type Param = Vec<F>;
    type Output = F;
    type Hessian = ();
    type Jacobian = ();
    type Float = F;

    fn apply(&self, params: &Self::Param) -> Result<Self::Output, anyhow::Error> {
        let distribution = D::try_from(params)?;
        Ok(-self.input_state.log_likelihood(&distribution))
    }
}

pub trait Fitter<S, P> {
    fn fit(&self) -> Result<P, Error>;
}

macro_rules! impl_fit {
    ($ty: ty) => {
        impl<S, D> Fitter<S, D> for BaseFitter<S, D, $ty>
        where
            S: LogLikelihood<D, $ty> + InitialSolvePoint<D>,
            D: for<'a> TryFrom<&'a [$ty], Error = Error> + Into<Vec<$ty>> + Debug,
        {
            fn fit(&self) -> Result<D, Error> {
                let initial_point: Vec<$ty> = self.input_state.initial_solve_point().into();
                let initial_simplex = initial_point.initial_simplex()?;

                let solver = NelderMead::new().with_initial_params(initial_simplex);

                let res = Executor::new(self, solver, initial_point.clone())
                    .max_iters(self.max_iterations)
                    .run()?;

                D::try_from(&res.state.best_param)
            }
        }
    };
}

impl_fit!(f32);
impl_fit!(f64);

#[cfg(test)]
mod tests {
    use super::*;
    use crate::distribution::weibull::WeibullDistribution;
    use crate::sample::{IntervalCensored, Weighted};
    use ndarray::prelude::*;

    const TOLERANCE: f64 = 1e-5;

    #[test]
    fn test_fit() {
        let data = Weighted {
            time: IntervalCensored {
                start: array![0., 2., 5., 10.],
                stop: array![2., 5., 10., 1e10f64],
            },
            weight: array![1000. - 376., 376. - 82., 82. - 7., 7.],
        };

        let fitter = BaseFitter::new(data);

        let actual: WeibullDistribution<_> = fitter.fit().unwrap();
        let expected = WeibullDistribution {
            scale: 2.0410538960706726,
            shape: 1.0170410407859767,
        };

        assert!((actual.scale - expected.scale).abs() < TOLERANCE);
        assert!((actual.shape - expected.shape).abs() < TOLERANCE);
    }
}