rsrl 0.8.1

A fast, extensible reinforcement learning framework in Rust
Documentation 
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

use crate::{
    policies::{Greedy, Policy, Random},
    Enumerable,
    Function,
};
use rand::Rng;

#[derive(Clone, Debug, Parameterised)]
#[cfg_attr(
    feature = "serde",
    derive(Serialize, Deserialize),
    serde(crate = "serde_crate")
)]
pub struct EpsilonGreedy<Q> {
    #[weights]
    greedy: Greedy<Q>,
    random: Random,

    pub epsilon: f64,
}

impl<Q> EpsilonGreedy<Q> {
    pub fn new(greedy: Greedy<Q>, random: Random, epsilon: f64) -> Self {
        EpsilonGreedy {
            greedy,
            random,

            epsilon,
        }
    }
}

impl<S, Q> Function<(S,)> for EpsilonGreedy<Q>
where Q: Enumerable<(S,), Output = Vec<f64>>
{
    type Output = Vec<f64>;

    fn evaluate(&self, (s,): (S,)) -> Vec<f64> {
        let prs = self.greedy.evaluate((s,));
        let pr = self.epsilon / prs.len() as f64;

        prs.into_iter()
            .map(|p| pr + p * (1.0 - self.epsilon))
            .collect()
    }
}

impl<S, A, Q> Function<(S, A)> for EpsilonGreedy<Q>
where
    A: std::borrow::Borrow<usize>,
    Q: Enumerable<(S,), Output = Vec<f64>>,
{
    type Output = f64;

    fn evaluate(&self, (s, a): (S, A)) -> f64 {
        let prs = self.greedy.evaluate((s,));
        let pr = self.epsilon / prs.len() as f64;

        pr + (1.0 - self.epsilon) * prs[*a.borrow()]
    }
}

impl<S, Q> Enumerable<(S,)> for EpsilonGreedy<Q>
where Q: Enumerable<(S,), Output = Vec<f64>>
{
    fn evaluate_index(&self, (s,): (S,), index: usize) -> f64 { self.evaluate((s, index)) }
}

impl<S, Q> Policy<S> for EpsilonGreedy<Q>
where Q: Enumerable<(S,), Output = Vec<f64>>
{
    type Action = usize;

    fn sample<R: Rng + ?Sized>(&self, rng: &mut R, s: S) -> usize {
        if rng.gen_bool(self.epsilon) {
            self.random.sample(rng, s)
        } else {
            self.greedy.sample(rng, s)
        }
    }

    fn mode(&self, s: S) -> usize { self.greedy.mode(s) }
}

#[cfg(test)]
mod tests {
    use approx::assert_abs_diff_eq;
    use crate::{
        fa::mocking::MockQ,
        policies::{EnumerablePolicy, EpsilonGreedy, Policy, Greedy, Random},
        Function,
    };
    use rand::thread_rng;

    #[test]
    fn test_sampling() {
        let mut rng = thread_rng();

        let q = MockQ::new_shared(Some(vec![1.0, 0.0].into()));
        let p = EpsilonGreedy::new(Greedy::new(q), Random::new(2), 0.5);

        let mut n0: f64 = 0.0;
        let mut n1: f64 = 0.0;
        for _ in 0..10000 {
            match p.sample(&mut rng, &vec![].into()) {
                0 => n0 += 1.0,
                _ => n1 += 1.0,
            }
        }

        assert!((0.75 - n0 / 10000.0).abs() < 0.05);
        assert!((0.25 - n1 / 10000.0).abs() < 0.05);
    }

    #[test]
    fn test_probabilites() {
        let q = MockQ::new_shared(None);
        let p = EpsilonGreedy::new(Greedy::new(q), Random::new(5), 0.5);

        p.evaluate((vec![1.0, 0.0, 0.0, 0.0, 0.0],))
            .into_iter()
            .zip([0.6, 0.1, 0.1, 0.1, 0.1].iter())
            .for_each(|(x, y)| assert_abs_diff_eq!(x, y, epsilon = 1e-6));

        p.evaluate((vec![0.0, 0.0, 0.0, 0.0, 1.0],))
            .into_iter()
            .zip([0.1, 0.1, 0.1, 0.1, 0.6].iter())
            .for_each(|(x, y)| assert_abs_diff_eq!(x, y, epsilon = 1e-6));

        p.evaluate((vec![1.0, 0.0, 0.0, 0.0, 1.0],))
            .into_iter()
            .zip([0.35, 0.1, 0.1, 0.1, 0.35].iter())
            .for_each(|(x, y)| assert_abs_diff_eq!(x, y, epsilon = 1e-6));
    }

    #[test]
    fn test_probabilites_uniform() {
        let q = MockQ::new_shared(Some(vec![-1.0, 0.0, 0.0, 0.0].into()));
        let p = EpsilonGreedy::new(Greedy::new(q), Random::new(4), 1.0);

        p.evaluate((vec![-1.0, 0.0, 0.0, 0.0],))
            .into_iter()
            .zip([0.25, 0.25, 0.25, 0.25].iter())
            .for_each(|(x, y)| assert_abs_diff_eq!(x, y, epsilon = 1e-6));
    }
}