reduce initialization weights

ml-agents/mlagents/trainers/ppo/trainer.py

                 baseline_estimates=baseline_estimates,
                 v_estimates=v_estimates,
                 value_next=value_next[name],
-                gamma=1,
+                gamma=self.optimizer.reward_signals[name].gamma,
                 lambd=1,
             )
 
             #    gamma=self.optimizer.reward_signals[name].gamma,
             #    lambd=self.hyperparameters.lambd,
             #)
+            #print(local_rewards)
+            #print(test_v)
+            #print(value_next[name])
             self._stats_reporter.add_stat(
                 f"Policy/{self.optimizer.reward_signals[name].name.capitalize()} Sum Rewards",
                 np.mean(test_v),

ml-agents/mlagents/trainers/torch/networks.py

         self.linear_encoder = LinearEncoder(
             encoder_input_size, network_settings.num_layers, self.h_size
         )
+        self.linear_encoder = LinearEncoder(
+            encoder_input_size,
+            network_settings.num_layers,
+            self.h_size,
+            kernel_gain=(0.125 / self.h_size) ** 0.5,
+        )
+
 
         if self.use_lstm:
             self.lstm = LSTM(self.h_size, self.m_size)