discard outliers before fitting

1 files changed, 27 insertions, 2 deletions

diff --git a/src/crates/candelabra/src/cost/fit.rs b/src/crates/candelabra/src/cost/fit.rs
index a9262b9..a6dbce0 100644
--- a/src/crates/candelabra/src/cost/fit.rs
+++ b/src/crates/candelabra/src/cost/fit.rs
@@ -2,6 +2,7 @@
 //! Based on code from al-jshen: <https://github.com/al-jshen/compute/tree/master>
 
 use super::benchmark::Observation;
+use log::trace;
 use na::{Dyn, OVector};
 use nalgebra::{dimension, Matrix, VecStorage};
 use serde::{Deserialize, Serialize};
@@ -19,8 +20,8 @@ pub type Cost = f64;
 impl Estimator {
     /// Fit from the given set of observations, using the least squared method.
     pub fn fit(results: &[Observation]) -> Self {
-        // Shift down so that ys start at 0
-        let (xs, ys) = Self::to_data(results);
+        let results = Self::discard_outliers(results);
+        let (xs, ys) = Self::to_data(&results);
 
         let xv = Self::prepare_input(&xs);
         let xtx = xv.transpose() * xv.clone();
@@ -98,4 +99,28 @@ impl Estimator {
 
         (xs, ys)
     }
+
+    fn discard_outliers(results: &[Observation]) -> Vec<Observation> {
+        let mut ns = results.iter().map(|(n, _)| *n).collect::<Vec<_>>();
+        ns.dedup();
+        let mut new_results = Vec::with_capacity(results.len());
+        for &n in ns.iter() {
+            let mut n_results: Vec<_> = results.iter().filter(|(n2, _)| *n2 == n).collect();
+            let old_len = n_results.len();
+            n_results.sort_by(|(_, x1), (_, x2)| x1.partial_cmp(x2).unwrap());
+
+            let (_, lq) = n_results[n_results.len() / 4];
+            let (_, uq) = n_results[(n_results.len() / 2) + (n_results.len() / 4)];
+
+            n_results.retain(|(_, x)| *x >= *lq && *x < *uq);
+            trace!(
+                "Discarded {} outliers for n = {n}",
+                old_len - n_results.len()
+            );
+
+            new_results.extend(n_results);
+        }
+
+        new_results
+    }
 }