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
157
158
|
use std::collections::{HashMap, HashSet};
use rand::{self, seq::IndexedRandom};
use serde::{Deserialize, Serialize};
use crate::{config::Config, trajectory::Action};
fn compute_weight(past: &HashSet<String>, now: &HashSet<&String>) -> f32 {
let inter = now.iter().filter(|&&p| past.contains(p)).count();
if inter == 0 {
return 0.0;
}
(inter as f32) / ((now.len() as f32).sqrt() * (past.len() as f32).sqrt())
}
fn softmax_sample(items: &[(String, f32)], temperature: f32) -> Option<(String, f32)> {
if items.is_empty() {
return None;
}
let mut rng = rand::rng();
let max_val = items
.iter()
.map(|(_, v)| *v)
.fold(f32::NEG_INFINITY, f32::max);
let exp_values: Vec<_> = items
.iter()
.map(|(k, val)| (k, ((val - max_val) / temperature).exp()))
.collect();
let sum: f32 = exp_values.iter().map(|(_, v)| *v).sum();
let norm_values: Vec<_> = exp_values.iter().map(|&(k, v)| (k, v / sum)).collect();
norm_values
.choose_weighted(&mut rng, |item| item.1)
.map(|p| (p.0.clone(), p.1))
.ok()
}
#[derive(Serialize, Deserialize)]
enum Episode {
Continue {
group: HashSet<String>,
avoid: Option<String>,
committed: String,
},
Escape {
from: HashSet<String>,
to: String,
committed: String,
},
}
#[derive(Serialize, Deserialize, Default)]
pub struct Learner {
history: Vec<Episode>,
}
pub enum Learning {
SkipToMore(Vec<String>, String),
MoreToSkip(Vec<String>, String),
}
impl Learner {
pub fn learn(&mut self, learning: &Learning) {
let stamp = chrono::Utc::now().format("%+").to_string();
match learning {
Learning::SkipToMore(trajectory, new) => {
self.history.push(Episode::Escape {
from: trajectory.iter().map(String::clone).collect(),
to: new.clone(),
committed: stamp,
});
}
Learning::MoreToSkip(trajectory, new) => {
self.history.push(Episode::Continue {
group: trajectory.iter().map(String::clone).collect(),
avoid: Some(new.clone()),
committed: stamp,
});
}
}
}
pub fn sample(
&self,
trajectory: &[String],
action: &Action,
candidates: &HashSet<String>,
config: &Config,
) -> Option<(String, f32)> {
let trajectory: HashSet<_> = trajectory.iter().collect();
let candidates: Vec<_> = candidates
.iter()
.filter(|c| !trajectory.contains(c))
.collect();
if candidates.is_empty() {
return None;
}
let mut items: HashMap<String, f32> =
candidates.into_iter().map(|c| (c.clone(), 0.0)).collect();
for episode in &self.history {
match (episode, action) {
(Episode::Escape { from, to, .. }, Action::Skip) => {
let w = compute_weight(from, &trajectory);
if let Some(to_w) = items.get_mut(to) {
*to_w += config.max_weight * w;
}
for f in from {
if let Some(from_weight) = items.get_mut(f) {
*from_weight -= config.mid_weight * w;
}
}
}
(Episode::Escape { from, to, .. }, Action::More) => {
let w = compute_weight(from, &trajectory);
if trajectory.contains(to) {
for f in from {
if let Some(f_w) = items.get_mut(f) {
*f_w -= config.low_weight * w;
}
}
}
}
(Episode::Continue { group, avoid, .. }, Action::More) => {
let w = compute_weight(group, &trajectory);
for g in group {
if let Some(g_w) = items.get_mut(g) {
*g_w += config.max_weight * w;
}
}
if let Some(a) = avoid
&& let Some(a_w) = items.get_mut(a)
{
*a_w -= config.low_weight * w;
}
}
(Episode::Continue { group, .. }, Action::Skip) => {
let w = compute_weight(group, &trajectory);
for g in group {
if let Some(v) = items.get_mut(g) {
*v -= config.high_weight * w;
}
}
}
}
}
let pairs: Vec<_> = items.into_iter().collect();
softmax_sample(&pairs, config.temperature)
}
}
|
