RLlib's offline dataset APIs enable working with experiences read from offline storage (e.g., disk, cloud storage, streaming systems, HDFS). For example, you might want to read experiences saved from previous training runs, or gathered from policies deployed in `web applications <https://arxiv.org/abs/1811.00260>`__. You can also log new agent experiences produced during online training for future use.
RLlib represents trajectory sequences (i.e., ``(s, a, r, s', ...)`` tuples) with `SampleBatch <https://github.com/ray-project/ray/blob/master/rllib/policy/sample_batch.py>`__ objects. Using a batch format enables efficient encoding and compression of experiences. During online training, RLlib uses `policy evaluation <rllib-concepts.html#policy-evaluation>`__ actors to generate batches of experiences in parallel using the current policy. RLlib also uses this same batch format for reading and writing experiences to offline storage.
In this example, we will save batches of experiences generated during online training to disk, and then leverage this saved data to train a policy offline using DQN. First, we run a simple policy gradient algorithm for 100k steps with ``"output": "/tmp/cartpole-out"`` to tell RLlib to write simulation outputs to the ``/tmp/cartpole-out`` directory.
The experiences will be saved in compressed JSON batch format:
..code-block:: text
$ ls -l /tmp/cartpole-out
total 11636
-rw-rw-r-- 1 eric eric 5022257 output-2019-01-01_15-58-57_worker-0_0.json
-rw-rw-r-- 1 eric eric 5002416 output-2019-01-01_15-59-22_worker-0_1.json
-rw-rw-r-- 1 eric eric 1881666 output-2019-01-01_15-59-47_worker-0_2.json
Then, we can tell DQN to train using these previously generated experiences with ``"input": "/tmp/cartpole-out"``. We disable exploration since it has no effect on the input:
In practice, when training on offline data, it is usually not straightforward to evaluate the trained policies using a simulator as in online RL. For example, in recommeder systems, rolling out a policy trained on offline data in a real-world environment can jeopardize your business if the policy is suboptimal. For these situations we can use `off-policy estimation <https://arxiv.org/abs/1911.06854>`__ methods which avoid the risk of evaluating a possibly sub-optimal policy in a real-world environment.
With RLlib's evaluation framework you can:
- Evaluate policies on a simulated environement, if available, using ``evaluation_config["input"] = "sampler"``. You can then monitor your policy's performance on tensorboard as it is getting trained (by using ``tensorboard --logdir=~/ray_results``).
- Use RLlib's off-policy estimation methods, which estimate the policy's performance on a separate offline dataset. To be able to use this feature, the evaluation dataset should contain ``action_prob`` key that represents the action probability distribution of the collected data so that we can do counterfactual evaluation.
RLlib supports the following off-policy estimators:
IS and WIS compute the ratio between the action probabilities under the behavior policy (from the dataset) and the target policy (the policy under evaluation), and use this ratio to estimate the policy's return. More details on this can be found in their respective papers.
DM and DR train a Q-model to compute the estimated return. By default, RLlib uses `Fitted-Q Evaluation (FQE) <https://arxiv.org/abs/1911.06854>`__ to train the Q-model. See `fqe_torch_model.py <https://github.com/ray-project/ray/blob/master/rllib/offline/estimators/fqe_torch_model.py>`__ for more details.
..note:: For a contextual bandit dataset, the ``dones`` key should always be set to ``True``. In this case, FQE reduces to fitting a reward model to the data.
RLlib's OPE estimators output six metrics:
-``v_behavior``: The discounted sum over rewards in the offline episode, averaged over episodes in the batch.
-``v_behavior_std``: The standard deviation corresponding to v_behavior.
-``v_target``: The OPE's estimated discounted return for the target policy, averaged over episodes in the batch.
-``v_target_std``: The standard deviation corresponding to v_target.
-``v_gain``: ``v_target / max(v_behavior, 1e-8)``, averaged over episodes in the batch. ``v_gain > 1.0`` indicates that the policy is better than the policy that generated the behavior data.
-``v_gain_std``: The standard deviation corresponding to v_gain.
As an example, we generate an evaluation dataset for off-policy estimation:
**Estimator Python API:** For greater control over the evaluation process, you can create off-policy estimators in your Python code and call ``estimator.train(batch)`` to perform any neccessary training and ``estimator.estimate(batch)`` to perform counterfactual estimation. The estimators take in an RLlib Policy object and gamma value for the environment, along with additional estimator-specific arguments (e.g. ``q_model_config`` for DM and DR). You can take a look at the example config parameters of the q_model_config `here <https://github.com/ray-project/ray/blob/master/rllib/offline/estimators/fqe_torch_model.py>`__. You can also write your own off-policy estimator by subclassing from the `OffPolicyEstimator <https://github.com/ray-project/ray/blob/master/rllib/offline/estimators/off_policy_estimator.py>`__ base class.
When the env does not support simulation (e.g., it is a web application), it is necessary to generate the ``*.json`` experience batch files outside of RLlib. This can be done by using the `JsonWriter <https://github.com/ray-project/ray/blob/master/rllib/offline/json_writer.py>`__ class to write out batches.
This `runnable example <https://github.com/ray-project/ray/blob/master/rllib/examples/saving_experiences.py>`__ shows how to generate and save experience batches for CartPole-v0 to disk:
However, for on-policy algorithms like PPO, you'll need to pass in the extra values added during policy evaluation and postprocessing to ``batch_builder.add_values()``, e.g., ``logits``, ``vf_preds``, ``value_target``, and ``advantages`` for PPO. This is needed since the calculation of these values depends on the parameters of the *behaviour* policy, which RLlib does not have access to in the offline setting (in online training, these values are automatically added during policy evaluation).
Note that for on-policy algorithms, you'll also have to throw away experiences generated by prior versions of the policy. This greatly reduces sample efficiency, which is typically undesirable for offline training, but can make sense for certain applications.
RLlib supports multiplexing inputs from multiple input sources, including simulation. For example, in the following example we read 40% of our experiences from ``/tmp/cartpole-out``, 30% from ``hdfs:/archive/cartpole``, and the last 30% is produced via policy evaluation. Input sources are multiplexed using `np.random.choice <https://docs.scipy.org/doc/numpy-1.15.0/reference/generated/numpy.random.choice.html>`__:
Similar to scaling online training, you can scale offline I/O throughput by increasing the number of RLlib workers via the ``num_workers`` config. Each worker accesses offline storage independently in parallel, for linear scaling of I/O throughput. Within each read worker, files are chosen in random order for reads, but file contents are read sequentially.
You can also define supervised model losses over offline data. This requires defining a `custom model loss <rllib-models.html#supervised-model-losses>`__. We provide a convenience function, ``InputReader.tf_input_ops()``, that can be used to convert any input reader to a TF input pipeline. For example:
..code-block:: python
def custom_loss(self, policy_loss):
input_reader = JsonReader("/tmp/cartpole-out")
# print(input_reader.next()) # if you want to access imperatively
See `custom_model_loss_and_metrics.py <https://github.com/ray-project/ray/blob/master/rllib/examples/custom_model_loss_and_metrics.py>`__ for a runnable example of using these TF input ops in a custom loss.
