big ram network

Author: sygi

deep_q_rl/ale_experiment.py

@@ -105,7 +105,7 @@ def _act(self, action):
         buffer
 
         """
-        reward = self.ale.act(action) + 1 # warning - this changes regular behaviour
+        reward = self.ale.act(action)
         index = self.buffer_count % self.buffer_length
 
         self.ale.getScreenGrayscale(self.screen_buffer[index, ...])

deep_q_rl/q_network.py

@@ -209,6 +209,9 @@ def build_network(self, network_type, input_width, input_height,
         elif network_type == "ram_dropout":
             return self.build_ram_dropout_network(input_width, input_height,
                     output_dim, num_frames, batch_size)
+        elif network_type == "big_ram":
+            return self.build_big_ram_network(input_width, input_height,
+                    output_dim, num_frames, batch_size)
         else:
             raise ValueError("Unrecognized network: {}".format(network_type))
 
@@ -487,6 +490,58 @@ def build_ram_network(self, input_width, input_height, output_dim,
 
         return l_out
 
+    def build_big_ram_network(self, input_width, input_height, output_dim,
+                          num_frames, batch_size):
+        """
+        Build a 5-layer network using only the information from the ram.
+        """
+
+        self.l_ram_in = lasagne.layers.InputLayer(
+            shape=(batch_size, self.RAM_SIZE)
+        )
+
+        l_hidden1 = lasagne.layers.DenseLayer(
+            self.l_ram_in,
+            num_units=self.RAM_SIZE,
+            nonlinearity=lasagne.nonlinearities.rectify,
+            W=lasagne.init.HeUniform(),
+            b=lasagne.init.Constant(.1)
+        )
+
+        l_hidden2 = lasagne.layers.DenseLayer(
+            l_hidden1,
+            num_units=self.RAM_SIZE,
+            nonlinearity=lasagne.nonlinearities.rectify,
+            W=lasagne.init.HeUniform(),
+            b=lasagne.init.Constant(.1)
+        )
+
+        l_hidden3 = lasagne.layers.DenseLayer(
+            l_hidden2,
+            num_units=self.RAM_SIZE,
+            nonlinearity=lasagne.nonlinearities.rectify,
+            W=lasagne.init.HeUniform(),
+            b=lasagne.init.Constant(.1)
+        )
+
+        l_hidden4 = lasagne.layers.DenseLayer(
+            l_hidden3,
+            num_units=self.RAM_SIZE,
+            nonlinearity=lasagne.nonlinearities.rectify,
+            W=lasagne.init.HeUniform(),
+            b=lasagne.init.Constant(.1)
+        )
+
+        l_out = lasagne.layers.DenseLayer(
+            l_hidden4,
+            num_units=output_dim,
+            nonlinearity=None,
+            W=lasagne.init.HeUniform(),
+            b=lasagne.init.Constant(.1)
+        )
+
+        return l_out
+
     def build_ram_dropout_network(self, input_width, input_height, output_dim,
             num_frames, batch_size):
         """