This shows you the differences between two versions of the page.
cs501r_f2018:lab9 [2018/11/19 21:17] wingated |
cs501r_f2018:lab9 [2021/06/30 23:42] |
||
---|---|---|---|
Line 1: | Line 1: | ||
- | |||
- | ====Objective:==== | ||
- | |||
- | * To implement the Proximal Policy Optimization algorithm, and learn about the use of deep learning in the context of deep RL. | ||
- | |||
- | ---- | ||
- | ====Deliverable:==== | ||
- | |||
- | For this lab, you will turn in a colab notebook that implements the proximal policy optimization (PPO) algorithm. You must provide tangible proof that your algorithm is working. | ||
- | |||
- | ---- | ||
- | ====Grading standards:==== | ||
- | |||
- | Your notebook will be graded on the following: | ||
- | |||
- | * 45% Proper design, creation and debugging of an actor and critic networks | ||
- | * 25% Proper implementation of the PPO loss function and objective on cart-pole ("CartPole-v0") | ||
- | * 20% Implementation and demonstrated learning of PPO on another domain of your choice (**except** VizDoom) | ||
- | * 10% Visualization of policy return as a function of training | ||
- | |||
- | ---- | ||
- | ====Description:==== | ||
- | |||
- | For this lab, you will implement the PPO algorithm, and train it on a few simple worlds from the OpenAI gym test suite of problems. | ||
- | |||
- | You may use any code you want from the internet to help you understand how to implement this, but **all final code must be your own**. | ||
- | |||
- | To successfully complete this lab, you must do the following: | ||
- | |||
- | * Implement a policy network. This should be a mapping from states to probabilities over actions. | ||
- | * Implement a value network. This should be a mapping from states to the value of that state. | ||
- | * Implement a loss function for the value network. This should compare the estimated value to the observed return. | ||
- | * Implement the PPO loss function. | ||
- | * Train the value and policy networks. This will at least involve: | ||
- | - Generating data by generating roll-outs (according to the current policy) | ||
- | - Calculating the actual discounted sum of rewards for each rollout | ||
- | - Using the value network to estimate the value of each rollout | ||
- | - Using actual and estimated value to calculate the advantage | ||
- | |||
- | Part of this lab involves a demonstration that your implementation is working properly. Most likely, this will be a measure of return / reward / value over time; as the policy improves, we should see it increase. | ||
- | |||
- | Some domains have other, more natural measures of performance. For example, for the cart-pole problem, we can measure how many iterations it takes before the pole falls over and we reach a terminal state; as the policy improves, this number improves. | ||
- | |||
- | You are also welcome to be creative in your illustrations - you could use a video of the agent doing something reasonable, for example. | ||
- | |||
- | ---- | ||
- | ====Background and documentation:==== | ||
- | |||
- | Here is [[https://gym.openai.com/|the OpenAI gym worlds]] | ||
- | |||
- | Here is a [[https://blog.openai.com/openai-baselines-ppo/ | blog post introducing the idea]]. | ||
- | |||
- | Here is the [[https://arxiv.org/pdf/1707.06347.pdf | paper with a technical description of the algorithm ]]. | ||
- | |||
- | Here is a [[https://www.youtube.com/watch?v=5P7I-xPq8u8 | video describing it at a high level ]]. | ||
- | |||
- | ---- | ||
- | ====Hints and helps:==== | ||
- | |||
- | **Update**: Here is our | ||
- | [[https://github.com/joshgreaves/reinforcement-learning|our lab's implementation of PPO]]. NOTE: because this code comes with a complete implementation of running on VizDoom, **you may not use that as your additional test domain.** | ||
- | |||
- | |||
- | Here is some code from our reference implementation. Hopefully it will serve as a good outline of what you need to do. | ||
- | |||
- | <code python> | ||
- | |||
- | import gym | ||
- | import torch | ||
- | ... | ||
- | |||
- | class PolicyNetwork(nn.Module): | ||
- | ... | ||
- | |||
- | |||
- | class ValueNetwork(nn.Module): | ||
- | .... | ||
- | |||
- | class AdvantageDataset(Dataset): | ||
- | def __init__(self, experience): | ||
- | super(AdvantageDataset, self).__init__() | ||
- | self._exp = experience | ||
- | self._num_runs = len(experience) | ||
- | self._length = reduce(lambda acc, x: acc + len(x), experience, 0) | ||
- | | ||
- | def __getitem__(self, index): | ||
- | idx = 0 | ||
- | seen_data = 0 | ||
- | current_exp = self._exp[0] | ||
- | while seen_data + len(current_exp) - 1 < index: | ||
- | seen_data += len(current_exp) | ||
- | idx += 1 | ||
- | current_exp = self._exp[idx] | ||
- | chosen_exp = current_exp[index - seen_data] | ||
- | return chosen_exp[0], chosen_exp[4] | ||
- | | ||
- | def __len__(self): | ||
- | return self._length | ||
- | | ||
- | | ||
- | class PolicyDataset(Dataset): | ||
- | def __init__(self, experience): | ||
- | super(PolicyDataset, self).__init__() | ||
- | self._exp = experience | ||
- | self._num_runs = len(experience) | ||
- | self._length = reduce(lambda acc, x: acc + len(x), experience, 0) | ||
- | | ||
- | def __getitem__(self, index): | ||
- | idx = 0 | ||
- | seen_data = 0 | ||
- | current_exp = self._exp[0] | ||
- | while seen_data + len(current_exp) - 1 < index: | ||
- | seen_data += len(current_exp) | ||
- | idx += 1 | ||
- | current_exp = self._exp[idx] | ||
- | chosen_exp = current_exp[index - seen_data] | ||
- | return chosen_exp | ||
- | | ||
- | def __len__(self): | ||
- | return self._length | ||
- | |||
- | |||
- | def main(): | ||
- | |||
- | env = gym.make('CartPole-v0') | ||
- | policy = PolicyNetwork(4, 2) | ||
- | value = ValueNetwork(4) | ||
- | |||
- | policy_optim = optim.Adam(policy.parameters(), lr=1e-2, weight_decay=0.01) | ||
- | value_optim = optim.Adam(value.parameters(), lr=1e-3, weight_decay=1) | ||
- | | ||
- | # ... more stuff here... | ||
- | |||
- | # Hyperparameters | ||
- | epochs = 1000 | ||
- | env_samples = 100 | ||
- | episode_length = 200 | ||
- | gamma = 0.9 | ||
- | value_epochs = 2 | ||
- | policy_epochs = 5 | ||
- | batch_size = 32 | ||
- | policy_batch_size = 256 | ||
- | epsilon = 0.2 | ||
- | |||
- | for _ in range(epochs): | ||
- | # generate rollouts | ||
- | rollouts = [] | ||
- | for _ in range(env_samples): | ||
- | # don't forget to reset the environment at the beginning of each episode! | ||
- | # rollout for a certain number of steps! | ||
- | |||
- | print('avg standing time:', standing_len / env_samples) | ||
- | calculate_returns(rollouts, gamma) | ||
- | |||
- | # Approximate the value function | ||
- | value_dataset = AdvantageDataset(rollouts) | ||
- | value_loader = DataLoader(value_dataset, batch_size=batch_size, shuffle=True, pin_memory=True) | ||
- | for _ in range(value_epochs): | ||
- | # train value network | ||
- | |||
- | calculate_advantages(rollouts, value) | ||
- | |||
- | # Learn a policy | ||
- | policy_dataset = PolicyDataset(rollouts) | ||
- | policy_loader = DataLoader(policy_dataset, batch_size=policy_batch_size, shuffle=True, pin_memory=True) | ||
- | for _ in range(policy_epochs): | ||
- | # train policy network | ||
- | |||
- | |||
- | |||
- | </code> | ||