haoyifan add code

code/README.txt

+# Environment
+A speaker-listener referential game based on reinforcement learning algorithm
+
+# Agents (Listener and Speaker)
+Stochastic Policy Gradient agents without parameter sharing or network connecting
+
+# Code structure
+'Agent_algorithm.py': contains code for the whole referential game framework
+    1). class Speaker(): algorithm and structure of the speaker
+    2). class Listener(): algorithm and structure of the listener
+    3). main(): the top function of all code, including settings, running process and evaluation of the referential game
+
+'metrics.py': contrains code for getting the probability distribution about symbols and concepts, and for computing the MIS, which is a metric to measure compositionality in our paper
+    1). update_speaker_prob(): getting policy and probability distribution of the speaker
+    2). update_listener_prob(): getting policy and probability distribution of the listener
+    3). update_R_and_MIS(): getting the metric MIS
+    4). update_metric(): the top function of 'metrics.py'
+
+# Run
+python Agent_algorithm.py GPU_ID
+for example, if you want use GPU 0,1,2, you can run like: python Agent_algorithm 0,1,2
+
+# Logs
+run_logs/log_XXX: contains policies of agents during the training process and the emergent language after trainig
+result_logs/log_XXX: contains mutual information matrix M and the metric MIS