implement MLE ranker

This commit is contained in:
Ilion Beyst 2022-06-04 15:08:45 +02:00
parent ef19e3a9e7
commit f899fba8ad

View file

@ -2,14 +2,14 @@ use crate::{db::bots::Bot, DbPool};
use crate::db;
use crate::modules::matches::RunMatch;
use diesel::{PgConnection, QueryResult};
use rand::seq::SliceRandom;
use std::time::Duration;
use std::collections::HashMap;
use std::mem;
use std::time::{Duration, Instant};
use tokio;
const RANKER_INTERVAL: u64 = 60;
const START_RATING: f64 = 0.0;
const SCALE: f64 = 100.0;
const MAX_UPDATE: f64 = 0.1;
pub async fn run_ranker(db_pool: DbPool) {
// TODO: make this configurable
@ -31,6 +31,7 @@ pub async fn run_ranker(db_pool: DbPool) {
bots.choose_multiple(&mut rng, 2).cloned().collect()
};
play_ranking_match(selected_bots, db_pool.clone()).await;
recalculate_ratings(&db_conn).expect("could not recalculate ratings");
}
}
@ -43,45 +44,193 @@ async fn play_ranking_match(selected_bots: Vec<Bot>, db_pool: DbPool) {
code_bundles.push(code_bundle);
}
let code_bundle_refs = code_bundles.iter().map(|b| b).collect::<Vec<_>>();
let code_bundle_refs = code_bundles.iter().collect::<Vec<_>>();
let mut run_match = RunMatch::from_players(code_bundle_refs);
run_match
.store_in_database(&db_conn)
.expect("could not store match in db");
let outcome = run_match
run_match
.spawn(db_pool.clone())
.await
.expect("running match failed");
}
let mut ratings = Vec::new();
for bot in &selected_bots {
let rating = db::ratings::get_rating(bot.id, &db_conn)
.expect("could not get bot rating")
.unwrap_or(START_RATING);
ratings.push(rating);
fn recalculate_ratings(db_conn: &PgConnection) -> QueryResult<()> {
let start = Instant::now();
let match_stats = fetch_match_stats(db_conn)?;
let ratings = estimate_ratings_from_stats(match_stats);
for (bot_id, rating) in ratings {
db::ratings::set_rating(bot_id, rating, db_conn).expect("could not update bot rating");
}
let elapsed = Instant::now() - start;
println!("computed ratings in {} ms", elapsed.subsec_millis());
Ok(())
}
#[derive(Default)]
struct MatchStats {
total_score: f64,
num_matches: usize,
}
fn fetch_match_stats(db_conn: &PgConnection) -> QueryResult<HashMap<(i32, i32), MatchStats>> {
let matches = db::matches::list_matches(db_conn)?;
let mut match_stats = HashMap::<(i32, i32), MatchStats>::new();
for m in matches {
if m.match_players.len() != 2 {
continue;
}
let (mut a_id, mut b_id) = match (&m.match_players[0].bot, &m.match_players[1].bot) {
(Some(ref a), Some(ref b)) => (a.id, b.id),
_ => continue,
};
// score of player a
let mut score = match m.base.winner {
None => 0.5,
Some(0) => 1.0,
Some(1) => 0.0,
_ => panic!("invalid winner"),
};
// put players in canonical order: smallest id first
if b_id < a_id {
mem::swap(&mut a_id, &mut b_id);
score = 1.0 - score;
}
let entry = match_stats.entry((a_id, b_id)).or_default();
entry.num_matches += 1;
entry.total_score += score;
}
Ok(match_stats)
}
/// Tokenizes player ids to a set of consecutive numbers
struct PlayerTokenizer {
id_to_ix: HashMap<i32, usize>,
ids: Vec<i32>,
}
impl PlayerTokenizer {
fn new() -> Self {
PlayerTokenizer {
id_to_ix: HashMap::new(),
ids: Vec::new(),
}
}
// simple elo rating
let scores = match outcome.winner {
None => vec![0.5; 2],
Some(player_num) => {
// TODO: please get rid of this offset
let player_ix = player_num - 1;
let mut scores = vec![0.0; 2];
scores[player_ix] = 1.0;
scores
fn tokenize(&mut self, id: i32) -> usize {
match self.id_to_ix.get(&id) {
Some(&ix) => ix,
None => {
let ix = self.ids.len();
self.ids.push(id);
self.id_to_ix.insert(id, ix);
ix
}
}
};
}
for i in 0..2 {
let j = 1 - i;
fn detokenize(&self, ix: usize) -> i32 {
self.ids[ix]
}
let scaled_difference = (ratings[j] - ratings[i]) / SCALE;
let expected = 1.0 / (1.0 + 10f64.powf(scaled_difference));
let new_rating = ratings[i] + MAX_UPDATE * (scores[i] - expected);
db::ratings::set_rating(selected_bots[i].id, new_rating, &db_conn)
.expect("could not update bot rating");
fn player_count(&self) -> usize {
self.ids.len()
}
}
fn sigmoid(logit: f64) -> f64 {
1.0 / (1.0 + (-logit).exp())
}
fn estimate_ratings_from_stats(match_stats: HashMap<(i32, i32), MatchStats>) -> Vec<(i32, f64)> {
// map player ids to player indexes in the ratings array
let mut input_records = Vec::<RatingInputRecord>::with_capacity(match_stats.len());
let mut player_tokenizer = PlayerTokenizer::new();
for ((a_id, b_id), stats) in match_stats.into_iter() {
input_records.push(RatingInputRecord {
p1_ix: player_tokenizer.tokenize(a_id),
p2_ix: player_tokenizer.tokenize(b_id),
score: stats.total_score / stats.num_matches as f64,
})
}
let mut ratings = vec![0f64; player_tokenizer.player_count()];
optimize_ratings(&mut ratings, &input_records);
ratings
.into_iter()
.enumerate()
.map(|(ix, rating)| {
(
player_tokenizer.detokenize(ix),
rating * 100f64 / 10f64.ln(),
)
})
.collect()
}
struct RatingInputRecord {
/// index of first player
p1_ix: usize,
/// index of secord player
p2_ix: usize,
/// score of player 1 (= 1 - score of player 2)
score: f64,
}
fn optimize_ratings(ratings: &mut [f64], input_records: &[RatingInputRecord]) {
// TODO: group this in a params struct
let tolerance = 10f64.powi(-6);
let learning_rate = 0.1;
let max_iterations = 10000;
for _iteration in 0..max_iterations {
let mut gradients = vec![0f64; ratings.len()];
// calculate gradients
for record in input_records.iter() {
let predicted = sigmoid(ratings[record.p1_ix] - ratings[record.p2_ix]);
let gradient = predicted - record.score;
gradients[record.p1_ix] += gradient;
gradients[record.p2_ix] -= gradient;
}
// apply update step
let mut converged = true;
for (rating, gradient) in ratings.iter_mut().zip(&gradients) {
let update = learning_rate * gradient / input_records.len() as f64;
if update > tolerance {
converged = false;
}
*rating -= update;
}
if converged {
break;
}
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn test_optimize_ratings() {
let input_records = vec![RatingInputRecord {
p1_ix: 0,
p2_ix: 1,
score: 0.8,
}];
let mut ratings = vec![0.0; 2];
optimize_ratings(&mut ratings, &input_records);
assert!(sigmoid(ratings[0] - ratings[1]) - 0.8 < 10f64.powi(-6));
}
}