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

//! Exploration strategies of IQN.
use serde::{Deserialize, Serialize};
use std::default::Default;
use tch::Tensor;

#[allow(clippy::upper_case_acronyms)]
#[derive(Debug, Deserialize, Serialize, PartialEq, Clone)]
/// Explorers for IQN.
pub enum IqnExplorer {
    // /// Softmax action selection.
    // Softmax(Softmax),
    /// Epsilon-greedy action selection.
    EpsilonGreedy(EpsilonGreedy),
}

// /// Softmax explorer for IQN.
// pub struct Softmax {}

// #[allow(clippy::new_without_default)]
// impl Softmax {
//     /// Constructs softmax explorer.
//     pub fn new() -> Self { Self {} }

//     /// Takes an action based on the observation and the critic.
//     pub fn action<M>(&mut self, qnet: &M, obs: &Tensor) -> Tensor where
//         M: Model1<Input=Tensor, Output=Tensor>,
//     {
//         let a = qnet.forward(obs);
//         a.softmax(-1, tch::Kind::Float).multinomial(1, true)
//     }
// }

#[derive(Debug, Deserialize, Serialize, PartialEq, Clone)]
/// Epsilon-greedy explorer for IQN.
pub struct EpsilonGreedy {
    n_opts: usize,
    eps_start: f64,
    eps_final: f64,
    final_step: usize,
}

impl Default for EpsilonGreedy {
    fn default() -> Self {
        Self {
            n_opts: 0,
            eps_start: 1.0,
            eps_final: 0.02,
            final_step: 100_000,
        }
    }
}

#[allow(clippy::new_without_default)]
impl EpsilonGreedy {
    /// Constructs epsilon-greedy explorer.
    pub fn with_params(eps_start: f64, eps_final: f64, final_step: usize) -> IqnExplorer {
        IqnExplorer::EpsilonGreedy(Self {
            n_opts: 0,
            eps_start,
            eps_final,
            final_step,
        })
    }

    /// Constructs epsilon-greedy explorer.
    ///
    /// TODO: improve interface.
    pub fn with_final_step(final_step: usize) -> IqnExplorer {
        IqnExplorer::EpsilonGreedy(Self {
            n_opts: 0,
            eps_start: 1.0,
            eps_final: 0.02,
            final_step,
        })
    }

    /// Takes an action based on the observation and the critic.
    pub fn action(&mut self, action_value: Tensor) -> Tensor {
        let d = (self.eps_start - self.eps_final) / (self.final_step as f64);
        let eps = (self.eps_start - d * self.n_opts as f64).max(self.eps_final);
        let r = fastrand::f64();
        let is_random = r < eps;
        self.n_opts += 1;

        if is_random {
            let batch_size = action_value.size()[0];
            let n_actions = action_value.size()[1] as u32;
            Tensor::of_slice(
                (0..batch_size)
                    .map(|_| fastrand::u32(..n_actions) as i32)
                    .collect::<Vec<_>>()
                    .as_slice(),
            )
        } else {
            action_value.argmax(-1, true)
        }
    }
}