From 2483f393d9a3740b35606ca6acb6cb2df8ffdcd2 Mon Sep 17 00:00:00 2001
From: Florian Jung <flo@windfisch.org>
Date: Fri, 8 Jan 2016 18:07:19 +0100
Subject: fixed neuronal network approach:

- make input values fit into 0..1 range (not 0..5/0..7)
- reward of 0.5 instead of 10
- randomize initial weights much more to avoid degeneration
---
 sol.py | 23 ++++++++++++++---------
 1 file changed, 14 insertions(+), 9 deletions(-)

(limited to 'sol.py')

diff --git a/sol.py b/sol.py
index c01d7e1..e369b79 100644
--- a/sol.py
+++ b/sol.py
@@ -280,32 +280,36 @@ class QNN (QCommon):
         self.dumbtraining = False
         connection_rate = 1
         num_input = 2
-        #hidden = (20,20)
-        hidden = (20,10,7)
+        #hidden = (40,40)
+        hidden = (50,)
+        #hidden = (20,10,7)
         num_output = 4
         learning_rate = 0.7
 
         self.NN = libfann.neural_net()
         #self.NN.set_training_algorithm(libfann.TRAIN_BATCH)
-        self.NN.set_training_algorithm(libfann.TRAIN_RPROP)
+        #self.NN.set_training_algorithm(libfann.TRAIN_RPROP)
         #self.NN.set_training_algorithm(libfann.TRAIN_QUICKPROP)
         self.NN.create_sparse_array(connection_rate, (num_input,)+hidden+(num_output,))
+        self.NN.randomize_weights(-1,1)
         self.NN.set_learning_rate(learning_rate)
         self.NN.set_activation_function_hidden(libfann.SIGMOID_SYMMETRIC_STEPWISE)
         self.NN.set_activation_function_output(libfann.SIGMOID_SYMMETRIC_STEPWISE)
         #self.NN.set_activation_function_output(libfann.LINEAR)
     
     def eval(self,x,y = None):
+        #print ("eval "+str(x)+", "+str(y))
         if y==None: x,y = x
 
-        return self.NN.run([x,y])
+        #print ("self.NN.run("+str([x/7.,y/5.])+")")
+        return self.NN.run([x/7.,y/5.])
     
     def change(self, s, action, diff):
         oldval = self.eval(s)
         newval = list(oldval) # copy list
         newval[action] += diff
 
-        self.NN.train(list(s), newval)
+        self.NN.train([s[0]/7.,s[1]/5.], newval)
 
     # learn a transition "from oldstate by action into newstate with reward `reward`"
     # this does not necessarily mean that the action is instantly trained into the function 
@@ -321,7 +325,8 @@ class QNN (QCommon):
             self.train_on_minibatch()
 
     def train_on_minibatch(self):
-        n = min(30, len(self.learnbuffer))
+        n = min(300, len(self.learnbuffer))
+        if n < 300: return
         minibatch = random.sample(self.learnbuffer, n)
 
         inputs = []
@@ -329,15 +334,15 @@ class QNN (QCommon):
         for oldstate, action, newstate, reward in minibatch:
             diff, val = self.value_update(oldstate, action, newstate, reward)
             val[action] += diff
-            inputs += [ list(oldstate) ]
+            inputs += [ [oldstate[0]/7., oldstate[1]/5.] ]
             outputs += [ val ]
 
         #print("training minibatch of size %i:\n%s\n%s\n\n"%(n, str(inputs), str(outputs)))
 
         training_data = libfann.training_data()
         training_data.set_train_data(inputs, outputs)
-        #self.NN.train_epoch(training_data)
-        self.NN.train_on_data(training_data, 5, 0, 0)
+        self.NN.train_epoch(training_data)
+        #self.NN.train_on_data(training_data, 5, 0, 0)
         #print(".")
 
     # must be called on every end-of-episode. might trigger batch-training or whatever.
-- 
cgit v1.2.1