From 0b7941e5b0f0ca100f1ea9e7e2022721265b0cc6 Mon Sep 17 00:00:00 2001
From: VasilyShcherbinin <vasily.shcherbinin@outlook.com>
Date: Sat, 2 Mar 2019 12:26:23 +0000
Subject: [PATCH] High-Medium-Low compress
---
XCS/RESULTS_FILE.txt | 7 +++++++
XCS/XCS_ClassifierSet.py | 2 ++
XCS/XCS_Configuration_File.txt | 4 ++--
3 files changed, 11 insertions(+), 2 deletions(-)
diff --git a/XCS/RESULTS_FILE.txt b/XCS/RESULTS_FILE.txt
index ae1a1cc..c6011ae 100644
--- a/XCS/RESULTS_FILE.txt
+++ b/XCS/RESULTS_FILE.txt
@@ -1,3 +1,10 @@
Rule Compaction XCS
I = 10000
+
+ Accuracy: 0.56494140625 Total time: 91.89 Rules: 6
+ Accuracy: 0.58984375 Total time: 92.94 Rules: 18
+ Accuracy: 0.6123046875 Total time: 92.56 Rules: 9
+ Accuracy: 0.56396484375 Total time: 92.41 Rules: 12
+ Accuracy: 0.60302734375 Total time: 92.17 Rules: 14
+ Accuracy: 0.6201171875 Total time: 92.56 Rules: 4
diff --git a/XCS/XCS_ClassifierSet.py b/XCS/XCS_ClassifierSet.py
index eeca4e4..6aab06f 100644
--- a/XCS/XCS_ClassifierSet.py
+++ b/XCS/XCS_ClassifierSet.py
@@ -510,6 +510,8 @@ class ClassifierSet:
while i < len(self.population):
cl = self.population[i]
if cl.action_cnt <= cons.theta_del or cl.error >= cons.e0: #error threshold e0
+ #if cl.error >= cons.e0: # error threshold e0 -> I am going to remove inaccurate classifiers, but leave the inexperienced.
+ # This gives better accuracy, but too many rules for "people comprehension, i.e. 80. 64%
self.micro_size -= cl.numerosity
self.population.pop(i)
else:
diff --git a/XCS/XCS_Configuration_File.txt b/XCS/XCS_Configuration_File.txt
index e9fda4b..aa6282a 100644
--- a/XCS/XCS_Configuration_File.txt
+++ b/XCS/XCS_Configuration_File.txt
@@ -27,7 +27,7 @@ trackingFrequency=0 # Specifies the number of iterations before each e
###### Learning Parameters - Generally just use default values.
######--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
nu=5 # (v) Power parameter used to determine the importance of high accuracy when calculating fitness. (typically set to 5, recommended setting of 1 in noisy data)
-chi=0.8 # (X) The probability of applying crossover in the GA. (typically set to 0.5-1.0)
+chi=0.8 # (X) The probability of applying crossover in the GA. (typically set to 0.5-1.0)
gamma=0 # Reinforcement Learning factor
phi=0 # If total prediction in matchset is less than phi times the mean prediction of population, covering occurs - not used.
upsilon=0.4 # (u) The probability of mutating an allele within an offspring.(typically set to 0.1-0.5)
@@ -55,7 +55,7 @@ doActionSetSubsumption=0 # Activate Subsumption? (1 is True, 0 is False).
selectionMethod=tournament # Select GA parent selection strategy ('tournament' or 'roulette')
differentParent=0 # Force different parents in tournament selection or not
theta_sel=0.5 # The fraction of the correct set to be included in tournament selection.
-crossoverMethod=uniform # option for type of crossover
+crossoverMethod=uniform # option for type of crossover
######--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
###### PopulationReboot - An option to begin e-LCS learning from an existing, saved rule population. Note that the training data is re-shuffled during a reboot.
--
GitLab