team-grasp / metalearning_rl

metalearning_RL

This project follows RL^2 and SNAIL paper and attempts to reproduce the results presented. rl2_train.py trains a model, given the arguments:

• num_workers (int): (Default: 1) This spawns specified amount of workers to perform parallel sampling (# of actors in PPO paper)
• model_type (str): (Default: gru) This chooses either gru (from RL^2) or snail as the model to train
• metalearn_epochs (int) (Default: 300) This is the number of iterations to train the model. After each iteration, a snapshot is taken for plotting the learning curve
• task (str): (Default: mdp) This supports either mdp (tabular MDP) or bandit (multi-armed bandit)
• learning_rate (float): (Default: 3e-4) This is the learning rate for the Adam optimizer
• gamma (float): (Default: 0.99) This is the discount factor
• num_actions (int): (Default: 5) This is the number of actions in the action space for the given task. This should only be tweaked for multi-armed bandit (unless new environment is provided)
• num_tasks (int): (Default: 1000) This specifies the number of tasks to learn from
• num_traj (int): (Default: 10) This specifies the number of trajectories to interact with given a task
• traj_len (int): (Default: 1) This specifies the length of the trajectory to sample from
• tau (float): (Default: 0.95) This is the GAE lambda parameter
• mini_batch_size (int): (Default: 256) This is the minibatch size M from PPO paper. This needs to be less than or equal to batch_size
• batch_size (int): (Default: 10000) This is the batch size T from the PPO paper. This essentially means we sample T actions before a PPO update
• ppo_epochs (int): (Default: 5) This is the PPO epoch K.
• clip_param (float): (Default: 0.1) This is the clipping factor epsilon from the PPO paper
• vf_coef (float): (Default: 0.5) This is the value function loss coefficient for the loss function
• ent_coef (float): (Default: 0.01) This is the entropy coefficient for the loss function
• max_grad_norm (float): (Default: 0.9) This clips the gradient if its norm exceeds the maximum norm allowed
• target_kl (float): (Default: 0.01) This is the mean KL that early stops the update if the KL divergence between old and new polcies is too high
• out_file (str): This is the name of the model file to be output

Installation:

You should be able to simply run pip install on the requirements.txt:
$ pip install -r requirements.txt

Sample Commands:

$ python rl2_train.py --out_file test_bandit.pt --batch_size 100 --num_tasks 100 --mini_batch_size 20 --num_traj 50 --tau 0.3 --gamma 0.99 --ppo_epochs 7 --learning_rate 0.005 --clip_param 0.2
$ python rl2_train.py --out_file test_10armed_bandit.pt --batch_size 100 --num_tasks 100 --mini_batch_size 20 --num_traj 50 --tau 0.3 --gamma 0.99 --ppo_epochs 7 --learning_rate 0.005 --clip_param 0.2 --num_actions 10
$ python rl2_train.py --out_file test_mdp_snail.pt --batch_size 100 --num_tasks 100 --mini_batch_size 20 --num_traj 50 --tau 0.3 --gamma 0.99 --ppo_epochs 7 --learning_rate 0.005 --clip_param 0.2 --model_type snail --metalearn_epochs 1500

References:

Papers:

• RL2: https://arxiv.org/abs/1611.02779
• SNAIL: https://arxiv.org/abs/1707.03141
• MAML: https://arxiv.org/abs/1703.03400
• PPO: https://arxiv.org/abs/1707.06347
• GAE: https://arxiv.org/abs/1506.02438

Codes:

• MAML-PyTorch: https://github.com/tristandeleu/pytorch-maml-rl
• RL-Adventure-2: https://github.com/higgsfield/RL-Adventure-2/blob/master/3.ppo.ipynb
• pytorch-ppo-acktr-a2c: https://github.com/ikostrikov/pytorch-a2c-ppo-acktr
• openAI baselines: https://github.com/openai/baselines/tree/master/baselines/ppo2
• openAI spinup: https://spinningup.openai.com/en/latest/_modules/spinup/algos/ppo/ppo.html
• snail-pytorch: https://github.com/eambutu/snail-pytorch
• rl2: https://github.com/VashishtMadhavan/rl2