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

// Copyright 2018 Stefan Kroboth
//
// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
// http://apache.org/licenses/LICENSE-2.0> or the MIT license <LICENSE-MIT or
// http://opensource.org/licenses/MIT>, at your option. This file may not be
// copied, modified, or distributed except according to those terms.

//! Backtracking Line Search
//!
//! TODO

use std;
use errors::*;
use ndarray::{Array1, arr1};
use ArgminSolver;
use result::ArgminResult;
use termination::TerminationReason;

/// Backtracking Line Search
pub struct BacktrackingLineSearch<'a> {
    /// Reference to cost function.
    cost_function: &'a Fn(&Array1<f64>) -> f64,
    /// Gradient
    gradient: &'a Fn(&Array1<f64>) -> Array1<f64>,
    /// Starting distance to the current point:
    alpha: f64,
    /// Maximum number of iterations
    max_iters: u64,
    /// Parameter `tau`
    tau: f64,
    /// Parameter `c`
    c: f64,
    /// Current state
    state: Option<BacktrackingLineSearchState>,
}

/// Current state of the backtracking line search algorithm
pub struct BacktrackingLineSearchState {
    /// Search direction
    p: Array1<f64>,
    /// Starting point
    x: Array1<f64>,
    /// Current cost value
    cost: f64,
    /// t (TODO)
    t: f64,
    /// Cost function value at starting point
    fx: f64,
    /// Current number of iteration
    iter: u64,
    /// Current alpha
    alpha: f64,
}

impl<'a> BacktrackingLineSearch<'a> {
    /// Initialize Backtracking Line Search
    ///
    /// Requires the cost function and gradient to be passed as parameter. The parameters
    /// `max_iters`, `tau`, and `c` are set to 100, 0.5 and 0.5, respectively.
    pub fn new(
        cost_function: &'a Fn(&Array1<f64>) -> f64,
        gradient: &'a Fn(&Array1<f64>) -> Array1<f64>,
    ) -> Self {
        BacktrackingLineSearch {
            cost_function: cost_function,
            gradient: gradient,
            alpha: 1.0,
            max_iters: 100,
            tau: 0.5,
            c: 0.5,
            state: None,
        }
    }

    /// Set the maximum distance from the starting point
    pub fn alpha(&mut self, alpha: f64) -> &mut Self {
        self.alpha = alpha;
        self
    }

    /// Set the maximum number of iterations
    pub fn max_iters(&mut self, max_iters: u64) -> &mut Self {
        self.max_iters = max_iters;
        self
    }

    /// Set c to a desired value between 0 and 1
    pub fn c(&mut self, c: f64) -> Result<&mut Self> {
        if c >= 1.0 || c <= 0.0 {
            return Err(ErrorKind::InvalidParameter(
                "BacktrackingLineSearch: Parameter `c` must satisfy 0 < c < 1.".into(),
            ).into());
        }
        self.c = c;
        Ok(self)
    }

    /// Set tau to a desired value between 0 and 1
    pub fn tau(&mut self, tau: f64) -> Result<&mut Self> {
        if tau >= 1.0 || tau <= 0.0 {
            return Err(ErrorKind::InvalidParameter(
                "BacktrackingLineSearch: Parameter `tau` must satisfy 0 < tau < 1.".into(),
            ).into());
        }
        self.tau = tau;
        Ok(self)
    }
}

impl<'a> ArgminSolver<'a> for BacktrackingLineSearch<'a> {
    type Parameter = Array1<f64>;
    type CostValue = f64;
    type Hessian = Array1<f64>;
    type StartingPoints = Array1<f64>;
    type ProblemDefinition = Array1<f64>;

    fn init(&mut self, p: Self::ProblemDefinition, x: &Self::StartingPoints) -> Result<()> {
        let m: f64 = p.t().dot(&((self.gradient)(x)));
        self.state = Some(BacktrackingLineSearchState {
            cost: std::f64::NAN,
            p: p,
            x: x.to_owned(),
            t: -self.c * m,
            fx: (self.cost_function)(x),
            iter: 0,
            alpha: self.alpha,
        });
        Ok(())
    }

    fn next_iter(&mut self) -> Result<ArgminResult<Self::Parameter, Self::CostValue>> {
        let mut state = self.state.take().unwrap();
        let param = &state.x + &(state.alpha * &state.p);
        state.cost = (self.cost_function)(&param);
        state.iter += 1;
        state.alpha *= self.tau;
        let mut out = ArgminResult::new(arr1(&[state.alpha]), std::f64::NAN, state.iter);
        self.state = Some(state);
        out.set_termination_reason(self.terminate());
        Ok(out)
    }

    /// Indicates whether any of the stopping criteria are met
    make_terminate!(self,
        self.state.as_ref().unwrap().iter >= self.max_iters, TerminationReason::MaxItersReached;
        self.state.as_ref().unwrap().fx - self.state.as_ref().unwrap().cost >= self.state.as_ref().unwrap().alpha * self.state.as_ref().unwrap().t, TerminationReason::TargetCostReached; 
    );

    /// Run gradient descent method
    make_run!(
        Self::ProblemDefinition,
        Self::StartingPoints,
        Self::Parameter,
        Self::CostValue
    );
}