Update best trials when state is updated

optuna_dashboard/ts/action.ts

@@ -23,6 +23,7 @@ import {
   artifactIsAvailable,
   reloadIntervalState,
 } from "./state"
+import { getDominatedTrials } from "./dominatedTrials"
 
 const localStorageGraphVisibility = "graphVisibility"
 const localStorageReloadInterval = "reloadInterval"
@@ -119,6 +120,34 @@ export const actionCreator = () => {
     newTrials[index] = newTrial
     const newStudy: StudyDetail = Object.assign({}, studyDetails[studyId])
     newStudy.trials = newTrials
+
+    // Update Best Trials
+    if (state === "Complete" && newStudy.directions.length === 1) {
+      // Single objective optimization
+      const bestValue = newStudy.best_trials.at(0)?.values?.at(0)
+      const currentValue = values?.at(0)
+      if (newStudy.best_trials.length === 0) {
+        newStudy.best_trials.push(newTrial)
+      } else if (bestValue !== undefined && currentValue !== undefined) {
+        if (newStudy.directions[0] === "minimize" && currentValue < bestValue) {
+          newStudy.best_trials = [newTrial]
+        } else if (
+          newStudy.directions[0] === "maximize" &&
+          currentValue > bestValue
+        ) {
+          newStudy.best_trials = [newTrial]
+        } else if (currentValue == bestValue) {
+          newStudy.best_trials.push(newTrial)
+        }
+      }
+    } else if (state === "Complete") {
+      // Multi objective optimization
+      newStudy.best_trials = getDominatedTrials(
+        newStudy.trials,
+        newStudy.directions
+      )
+    }
+
     setStudyDetailState(studyId, newStudy)
   }
 

optuna_dashboard/ts/dominatedTrials.ts

@@ -0,0 +1,40 @@
+const filterFunc = (trial: Trial, directions: StudyDirection[]): boolean => {
+  return (
+    trial.state === "Complete" &&
+    trial.values !== undefined &&
+    trial.values.length === directions.length &&
+    trial.values.every((v) => v !== "inf" && v !== "-inf")
+  )
+}
+
+export const getDominatedTrials = (
+  trials: Trial[],
+  directions: StudyDirection[]
+): Trial[] => {
+  // TODO(c-bata): Use log-linear algorithm like Optuna.
+  // TODO(c-bata): Use this function at GraphParetoFront.
+  const filteredTrials = trials.filter((t: Trial) => filterFunc(t, directions))
+
+  const normalizedValues: number[][] = []
+  filteredTrials.forEach((t) => {
+    if (t.values && t.values.length === directions.length) {
+      const trialValues = t.values.map((v, i) => {
+        return directions[i] === "minimize" ? (v as number) : (-v as number)
+      })
+      normalizedValues.push(trialValues)
+    }
+  })
+  const dominatedTrials: boolean[] = []
+  normalizedValues.forEach((values0: number[], i: number) => {
+    const dominated = normalizedValues.some((values1: number[], j: number) => {
+      if (i === j) {
+        return false
+      }
+      return values0.every((value0: number, k: number) => {
+        return values1[k] <= value0
+      })
+    })
+    dominatedTrials.push(dominated)
+  })
+  return filteredTrials.filter((_, i) => !dominatedTrials.at(i))
+}