Set new hyperparameter values for various strategies after more extensive training

Author: fjwillemsen

kernel_tuner/strategies/diff_evo.py

@@ -10,11 +10,11 @@
 from kernel_tuner.strategies.common import CostFunc
 
 _options = dict(
-    popsize=("population size", 50),
-    popsize_times_dimensions=("multiply population size with number of dimensions (True/False)", False),
+    popsize=("population size", 16),
+    popsize_times_dimensions=("multiply population size with number of dimensions (True/False)", True),
     maxiter=("maximum number of generations", int(1e15)),    # very large to avoid early stopping (stopping is managed by StopCriterionReached)
-    F=("mutation factor (differential weight)", 1.3),
-    CR=("crossover rate", 0.9),
+    F=("mutation factor (differential weight)", 0.7),
+    CR=("crossover rate", 0.6),
     method=("method", "best1bin"),
     constraint_aware=("constraint-aware optimization (True/False)", True),
 )

kernel_tuner/strategies/genetic_algorithm.py

@@ -10,10 +10,10 @@
 from kernel_tuner.strategies.common import CostFunc
 
 _options = dict(
-    popsize=("population size", 20),
-    maxiter=("maximum number of generations", 150),
+    popsize=("population size", 26),
+    maxiter=("maximum number of generations", 90),
     method=("crossover method to use, choose any from single_point, two_point, uniform, disruptive_uniform", "single_point"),
-    mutation_chance=("chance to mutate is 1 in mutation_chance", 5),
+    mutation_chance=("chance to mutate is 1 in mutation_chance", 55),
     constraint_aware=("constraint-aware optimization (True/False)", True),
 )
 

kernel_tuner/strategies/pso.py

@@ -11,11 +11,11 @@
 from kernel_tuner.strategies.common import CostFunc, scale_from_params
 
 _options = dict(
-    popsize=("Population size", 30),
-    maxiter=("Maximum number of iterations", 100),
+    popsize=("Population size", 50),
+    maxiter=("Maximum number of iterations", 190),
     w=("Inertia weight constant", 0.5),
-    c1=("Cognitive constant", 3.0),
-    c2=("Social constant", 0.5),
+    c1=("Cognitive constant", 3.5),
+    c2=("Social constant", 1.0),
     constraint_aware=("constraint-aware optimization (True/False)", True))
 
 def tune(searchspace: Searchspace, runner, tuning_options):

kernel_tuner/strategies/simulated_annealing.py

@@ -10,10 +10,10 @@
 from kernel_tuner.strategies.common import CostFunc
 
 
-_options = dict(T=("Starting temperature", 0.5),
-                T_min=("End temperature", 0.001),
+_options = dict(T=("Starting temperature", 0.1),
+                T_min=("End temperature", 0.0001),
                 alpha=("Alpha parameter", 0.9975),
-                maxiter=("Number of iterations within each annealing step", 2),
+                maxiter=("Number of iterations within each annealing step", 1),
                 constraint_aware=("constraint-aware optimization (True/False)", True))
 
 def tune(searchspace: Searchspace, runner, tuning_options):