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
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
|
use std::collections::{HashMap, HashSet};
use rand::{self, seq::IndexedRandom};
use serde::{Deserialize, Serialize};
use crate::{config::Config, trajectory::Action};
const DIM: usize = 16;
// -----------------------------------------------------------------------------
// helpers
fn dot(a: &[f32; DIM], b: &[f32; DIM]) -> f32 {
a.iter().zip(b).map(|(x, y)| x * y).sum()
}
fn norm(a: &[f32; DIM]) -> f32 {
dot(a, a).sqrt()
}
fn renorm(a: &mut [f32; DIM]) {
let n = norm(a).max(1e-6);
for x in a.iter_mut() {
*x /= n;
}
}
fn init_vec() -> [f32; DIM] {
let mut v = [0.0; DIM];
for x in v.iter_mut() {
*x = rand::random_range(-0.1..0.1);
}
renorm(&mut v);
v
}
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, Default)]
pub struct Learner {
embedding: HashMap<String, [f32; DIM]>,
}
impl Learner {
pub fn prune(&mut self, valid: &HashSet<String>) {
self.embedding.retain(|k, _v| valid.contains(k));
}
fn nudge(&mut self, target: &str, direction: &str, rate: f32) {
let t = *self
.embedding
.entry(direction.into())
.or_insert_with(init_vec);
let v = self.embedding.entry(target.into()).or_insert_with(init_vec);
for i in 0..DIM {
v[i] += rate * (t[i] - v[i]);
}
renorm(v);
}
}
// -----------------------------------------------------------------------------
// core
#[derive(Debug)]
pub enum Learning {
SkipExtend,
MoreExtend(Vec<String>, String),
SkipToMore(Vec<String>, String),
MoreToSkip(Vec<String>, String),
}
impl Learner {
pub fn learn(&mut self, config: &Config, learning: &Learning) {
match learning {
Learning::SkipExtend => {
/* This is the least informative case, it does not
follow that the trajectory contains similar or different
items, we may simply be seeking something in particular. */
}
Learning::MoreExtend(trajectory, new) => {
/* We're continuing a good run so `new` is similar to everything
in `trajectory`, but to account for the possibilty that our
mood has changed over the course of this streak we dampen
this feedback. */
for (distance, similar) in trajectory.iter().rev().enumerate() {
let damp = ((distance + 1) as f32).powf(-0.5);
self.nudge(new, similar, config.learning_rate * damp);
self.nudge(similar, new, config.learning_rate * damp * 0.5); // optional: symmetric, weaker
}
}
Learning::SkipToMore(trajectory, new) => {
/* We have learnt that `new` is different to everything in
`trajectory`, the strongest signal we have. */
for different in trajectory {
self.nudge(new, different, -config.learning_rate);
}
}
Learning::MoreToSkip(trajectory, new) => {
/* `new` could be different to everything in `trajectory`, or we
simply changed our minds, so we have only weak evidence of
difference. The positive coherence of trajectory was taken
care of during MoreExtend above. */
let damp = (trajectory.len() + 1) as f32;
for weakly_different in trajectory {
self.nudge(new, weakly_different, -config.learning_rate / damp);
}
}
}
}
fn mood(&self, trajectory: &[String]) -> Option<[f32; DIM]> {
let mut m = [0.0; DIM];
let mut any = false;
for (distance, historical) in trajectory.iter().rev().enumerate() {
if let Some(v) = self.embedding.get(historical) {
let damp = ((distance + 1) as f32).powf(-0.5);
for i in 0..DIM {
m[i] += damp * v[i];
}
any = true;
}
}
if !any || norm(&m) < 1e-6 {
return None;
}
renorm(&mut m);
Some(m)
}
pub fn sample(
&self,
trajectory: &[String],
action: &Action,
candidates: &HashSet<String>,
temperature: f32,
) -> Option<(String, f32)> {
let seen: HashSet<_> = trajectory.iter().collect();
let candidates: Vec<_> = candidates.iter().filter(|c| !seen.contains(c)).collect();
if candidates.is_empty() {
return None;
}
let Some(mood) = self.mood(trajectory) else {
let items: Vec<_> = candidates.iter().map(|&c| (c, 0.0)).collect();
return softmax_sample(&items, temperature);
};
let items: Vec<_> = candidates
.iter()
.map(|&c| {
let score = match self.embedding.get(c.as_str()) {
None => 0.0,
Some(v) => {
let cos = dot(v, &mood);
match action {
Action::More => cos,
Action::Skip => -cos.abs(),
}
}
};
(c, score)
})
.collect();
softmax_sample(&items, temperature)
}
}
|
