CommNet and BiCnet implementation in tensorflow
Train CommNet using DDPG algorithm
python train_comm_net.py
To find the optimal hyperparameters such as actor_lr or critic_lr, a simple grid search has been implemented. It launches multiple instances of the trainer in parallel based on the number of CPU cores.
python hypersearch.py
It is a simple game described in the BiCnet paper for testing if the communication works. The environment implements the crucial methods of the core gym interface from OpenAI
Each agent receives a scalar sampled between [−10, 10] under a truncated Gaussian. Each agent needs to output the sum of all inputs received among the agents. An agent gets a normalized reward between [0, 1] based on the absolute difference between the sum and its output.
