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:
**Off-policy estimation:** Since the input experiences are not from running simulations, RLlib cannot report the true policy performance during training. However, you can use ``tensorboard --logdir=~/ray_results`` to monitor training progress via other metrics such as estimated Q-value. Alternatively, `off-policy estimation <https://arxiv.org/pdf/1511.03722.pdf>`__ can be used, which requires both the source and target action probabilities to be available (i.e., the ``action_prob`` batch key). For DQN, this means enabling soft Q learning so that actions are sampled from a probability distribution:
This example plot shows the Q-value metric in addition to importance sampling (IS) and weighted importance sampling (WIS) gain estimates (>1.0 means there is an estimated improvement over the original policy):
**Estimator Python API:** For greater control over the evaluation process, you can create off-policy estimators in your Python code and call ``estimator.estimate(episode_batch)`` to perform counterfactual estimation as needed. The estimators take in a policy object and gamma value for the environment:
**Simulation-based estimation:** If true simulation is also possible (i.e., your env supports ``step()``), you can also set ``"input_evaluation": ["simulation"]`` to tell RLlib to run background simulations to estimate current policy performance. The output of these simulations will not be used for learning. Note that in all cases you still need to specify an environment object to define the action and observation spaces. However, you don't need to implement functions like reset() and step().
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:
RLlib assumes that input batches are of `postprocessed experiences <https://github.com/ray-project/ray/blob/cf21c634a390745ba6f8916b1f34f7b0453bc7dd/rllib/policy/policy.py#L376>`__. This isn't typically critical for off-policy algorithms (e.g., DQN's `post-processing <https://github.com/ray-project/ray/blob/cf21c634a390745ba6f8916b1f34f7b0453bc7dd/rllib/agents/dqn/dqn_tf_policy.py#L387>`__ is only needed if ``n_step > 1`` or ``worker_side_prioritization: True``). For off-policy algorithms, you can also safely set the ``postprocess_inputs: True`` config to auto-postprocess data.
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.
Mixing simulation and offline data
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
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.
