#random #markov #probability #chain #numeric

markovian

Simulation of Markov Processes as stochastic processes

5 releases

0.2.1 Jul 13, 2020
0.1.3 Apr 6, 2020
0.1.2 Apr 2, 2020
0.1.1 Mar 24, 2020
0.1.0 Mar 24, 2020

#1335 in Math

MIT/Apache

53KB
950 lines

markovian

Simulation of Stochastic processes.

Goal

Serve as an extension of the rand crate for sub-stochastic markovian processes.

Main features

  • Easy construction of Markov processes, including:
    • Discrete time
    • Continuous time (exponential clocks)
  • Type agnostic

Changelog

See Changelog.

Contribution

Your contribution is highly appreciated. Do not hesitate to open an issue or a pull request. Note that any contribution submitted for inclusion in the project will be licensed according to the terms of the dual license (MIT and Apache-2.0).


lib.rs:

Simulation of (sub-)stochastic processes.

Goal

Serve as an extension of the rand crate for sub-stochastic processes.

Examples

Discrete time

Construction of a random walk in the integers.

let init_state: i32 = 0;
let transition = |state: &i32| raw_dist![(0.5, state + 1), (0.5, state - 1)];
let rng = thread_rng();
let mut mc = markovian::MarkovChain::new(init_state, transition, rng);

Branching process

Construction using density p(0) = 0.3, p(1) = 0.4, p(2) = 0.3.

let init_state: u32 = 1;
let base_distribution = raw_dist![(0.3, 0), (0.4, 1), (0.3, 2)];
let rng = thread_rng();
let mut branching_process = markovian::BranchingProcess::new(init_state, base_distribution, rng);

Continuous time

Construction of a random walk in the integers, with expponential time for each transition.

let init_state: i32 = 0;
struct MyTransition;
impl markovian::Transition<i32, (f64, i32)> for MyTransition {
    fn sample_from<R: ?Sized>(&self, state: &i32, rng: &mut R) -> (f64, i32)
    where
        R: Rng
    {
        let time = Exp::new(2.0).unwrap().sample(rng);
        let step = Uniform::from(0..=1).sample(rng) * 2 - 1;
        (time, state + step)
    }
}
let transition = MyTransition;
let rng = thread_rng();
let mut mc = markovian::TimedMarkovChain::new(init_state, transition, rng);

Remarks

All methods are inline, by design.

Non-trivial ways to use the crate are described below, including time dependence, continuous space and non-markovian processes.

Time dependence

Include the time as part of the state of the process.

Examples

A random walk on the integers that tends to move more to the right as time goes by.

let init_state: (usize, i32) = (0, 0);
let transition = |(time, state): &(usize, i32)| raw_dist![
    (0.6 - 1.0 / (time + 2) as f64, (time + 1, state + 1)),
    (0.4 + 1.0 / (time + 2) as f64, (time + 1, state - 1))
];
let rng = thread_rng();
let mut mc = markovian::MarkovChain::new(init_state, &transition, rng);

// Take a sample of 10 elements 
mc.take(10).map(|(_, state)| state).collect::<Vec<i32>>();

Continuous space

Randomize the transition: return a random element together with a probability one

Examples

A random walk on the real line with variable step size.

let init_state: f64 = 0.0;
struct MyTransition;
impl markovian::Transition<f64, f64> for MyTransition {
    fn sample_from<R: ?Sized>(&self, state: &f64, rng: &mut R) -> f64
    where
        R: Rng
    {
        let step = Exp::new(2.0).unwrap().sample(rng);
        state + step
    }
}
let transition = MyTransition;
let rng = thread_rng();
let mut mc = markovian::MarkovChain::new(init_state, transition, rng);
mc.next();
 
// current_state is positive 
assert!(mc.state().unwrap() > &0.0);

Non markovian

Include history in the state. For example, instead of i32, use Vec<i32>.

Examples

A random walk on the integers that is atracted to zero in a non markovian fashion.

let init_state: Vec<i32> = vec![0];
let transition = |state: &Vec<i32>| {
    // New possible states
    let mut right = state.clone();
    right.push(state[state.len() - 1] + 1);
    let mut left = state.clone();
    left.push(state[state.len() - 1] - 1);

    // Some non markovian transtion
    let path_stadistic: i32 = state.iter().sum();
    if path_stadistic.is_positive() {
        raw_dist![
            (1.0 / (path_stadistic.abs() + 1) as f64, right), 
            (1.0 - 1.0 / (path_stadistic.abs() + 1) as f64, left)
        ]
    } else {
        raw_dist![
            (1.0 - 1.0 / (path_stadistic.abs() + 1) as f64, right), 
            (1.0 / (path_stadistic.abs() + 1) as f64, left)
        ]
    }
};
let rng = thread_rng();
let mut mc = markovian::MarkovChain::new(init_state, transition, rng);
 
// state has history
mc.next();
assert_eq!(mc.state().unwrap().len(), 2);

Dependencies

~1.1–1.7MB
~35K SLoC