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

use crate::{
    fa::{GradientUpdate, ScaledGradientUpdate, StateActionUpdate},
    params::*,
    policies::Policy,
    Differentiable,
    Function,
    Handler,
};
use ndarray::{ArrayBase, ArrayView2, Data, Ix2};
use rand::Rng;

#[derive(Clone, Debug, Parameterised)]
#[cfg_attr(
    feature = "serde",
    derive(Serialize, Deserialize),
    serde(crate = "serde_crate")
)]
pub struct Point<F>(pub F);

impl<F> Point<F> {
    pub fn new(fa: F) -> Self { Point(fa) }
}

impl<'s, S, A, F> Function<(S, A)> for Point<F>
where
    A: std::borrow::Borrow<F::Output>,
    F: Function<(S,)>,
    F::Output: PartialEq,
{
    type Output = f64;

    fn evaluate(&self, (s, a): (S, A)) -> f64 {
        let v = self.0.evaluate((s,));

        if v.eq(a.borrow()) {
            1.0
        } else {
            0.0
        }
    }
}

impl<'s, S, A, F> Differentiable<(S, A)> for Point<F>
where
    A: std::borrow::Borrow<<F as Function<(S,)>>::Output>,
    F: Function<(S,)> + Differentiable<(S, A)>,

    <F as Function<(S,)>>::Output: PartialEq,
{
    type Jacobian = F::Jacobian;

    fn grad(&self, (s, a): (S, A)) -> F::Jacobian { self.0.grad((s, a)) }

    fn grad_log(&self, (s, a): (S, A)) -> F::Jacobian { self.0.grad_log((s, a)) }
}

impl<S, F> Policy<S> for Point<F>
where
    F: Function<(S,)>,
    F::Output: PartialEq,
{
    type Action = F::Output;

    fn sample<R: Rng + ?Sized>(&self, _: &mut R, s: S) -> Self::Action { self.0.evaluate((s,)) }

    fn mode(&self, s: S) -> Self::Action { self.0.evaluate((s,)) }
}

impl<S, A, F> Handler<StateActionUpdate<S, A>> for Point<F>
where
    A: std::borrow::Borrow<f64>,
    F: for<'s> Function<(&'s S,), Output = f64> + Handler<StateActionUpdate<S, A>>,
{
    type Response = F::Response;
    type Error = F::Error;

    fn handle(&mut self, msg: StateActionUpdate<S, A>) -> Result<Self::Response, Self::Error> {
        let mode = self.0.evaluate((&msg.state,));
        let error = (msg.action.borrow() - mode) * msg.error;

        self.0.handle(StateActionUpdate {
            state: msg.state,
            action: msg.action,
            error,
        })
    }
}

impl<'m, D, F> Handler<GradientUpdate<&'m ArrayBase<D, Ix2>>> for Point<F>
where
    D: Data<Elem = f64>,
    F: Parameterised + Handler<GradientUpdate<ArrayView2<'m, f64>>>,
{
    type Response = F::Response;
    type Error = F::Error;

    fn handle(
        &mut self,
        msg: GradientUpdate<&'m ArrayBase<D, Ix2>>,
    ) -> Result<F::Response, F::Error>
    {
        let dim = self.0.weights_dim();

        self.0.handle(GradientUpdate(msg.0.slice(s![0..dim.0, 0..dim.1])))
    }
}

impl<'m, D, F> Handler<ScaledGradientUpdate<&'m ArrayBase<D, Ix2>>> for Point<F>
where
    D: Data<Elem = f64>,
    F: Parameterised + Handler<ScaledGradientUpdate<ArrayView2<'m, f64>>>,
{
    type Response = F::Response;
    type Error = F::Error;

    fn handle(
        &mut self,
        msg: ScaledGradientUpdate<&'m ArrayBase<D, Ix2>>,
    ) -> Result<F::Response, F::Error>
    {
        let dim = self.0.weights_dim();

        self.0.handle(ScaledGradientUpdate {
            alpha: msg.alpha,
            jacobian: msg.jacobian.slice(s![0..dim.0, 0..dim.1]),
        })
    }
}