add time to first because last is often naughty-listed; implement nau…

…ghty-list for nearest neighbour time predictor

bin/medley

@@ -26,7 +26,7 @@ def main():
         "-e",
         help="extra time to first? Add to last if not",
         type=bool,
-        default=False,
+        default=True,
     )
     global_parser.add_argument(
         "--kind",

medleysolver/timers.py

@@ -40,8 +40,11 @@ def __init__(self, init_lambda, confidence, T):
         self.init_lambda = init_lambda
         self.confidence = confidence
         self.T = T
+        self.naughtylist = set()
 
     def get_timeout(self, solver, times):
+        if solver in self.naughtylist:
+            return 0
         # want time based on times for same solver at nearby points
         timer = ExponentialDist(self.init_lambda, self.confidence, self.T)
         for (s, t) in times:
@@ -50,4 +53,6 @@ def get_timeout(self, solver, times):
         return timer.get_cutoff()
 
     def update(self, solver, time, timeout, success, error):
-        pass
+        assert(not success or not error)
+        if error:
+            self.naughtylist.add(solver)

runmedley.sh

@@ -1,5 +1,5 @@
-for seed in 0 1; do
-    for learner in thompson neighbor knearest random MLP linear exp3a exp3b exp3c; do
+for seed in 0; do #1; do
+    for learner in thompson knearest exp3a; do # neighbour random MLP linear exp3b exp3c; do
         if [ $learner = exp3a ]
         then
             learnconfig="exp3 --gamma 0.07"
@@ -12,10 +12,10 @@ for seed in 0 1; do
         else 
             learnconfig=$learner
         fi
-        for feature in both probes bow; do
-            for reward in binary bump exp; do
-                medley ./$1/ ./$1/${learner}_${feature}_${reward}_const_${seed}.csv   --classifier $learnconfig --seed $seed --feature_setting $feature --reward $reward --timeout_manager const --set_const 60
-                medley ./$1/ ./$1/${learner}_${feature}_${reward}_expo_${seed}.csv    --classifier $learnconfig --seed $seed --feature_setting $feature --reward $reward --timeout_manager expo
+        for feature in both; do #probes bow; do
+            for reward in bump; do # binary exp; do
+                # medley ./$1/ ./$1/${learner}_${feature}_${reward}_const_${seed}.csv   --classifier $learnconfig --seed $seed --feature_setting $feature --reward $reward --timeout_manager const --set_const 60
+                # medley ./$1/ ./$1/${learner}_${feature}_${reward}_expo_${seed}.csv    --classifier $learnconfig --seed $seed --feature_setting $feature --reward $reward --timeout_manager expo
                 medley ./$1/ ./$1/${learner}_${feature}_${reward}_nearest_${seed}.csv --classifier $learnconfig --seed $seed --feature_setting $feature --reward $reward --timeout_manager nearest --time_k 20
             done 
         done