Returns arr[start] + ... + arr[end]

Like forward() but with numpy arrays.

Returns arr[start] + ... + arr[end]

Warp frames to 84x84 as done in the Nature paper and later work.

Warp frames to 84x84 as done in the Nature paper and later work.

Warp frames to 84x84 as done in the Nature paper and later work.

Initialises the level.

Starts a new episode.

Save a model to a file, making sure that the file will never be partly written. This prevents the model from getting corrupted in the

Load training and test recordings from disk. Args: dirpaths: a sequence of directories where recordings are stored.

Load an array for the frame image at the given timestep index.

Yield `count` batches, where each batch is a list of (timestep, batch_idx) tuples.

Big obs takes a retro observation and an info dictionary and produces a higher resolution observation with the retro features tacked on.

A helper function to create a batch of environments. Args: num_envs: size of the batch. start: the starting worker index num

Create a single, wrapped environment. Args: idx: the ML-Agents worker index to use. clear: erase most of the info dictionary

Gets a random available position in a binary map array. Args: map_array: numpy array of the map to search an available position on. Returns:

Returns min(arr[start], ..., arr[end])

Bin reward to {+1, 0, -1} by its sign.

Build a Segment Tree data structure. https://en.wikipedia.org/wiki/Segment_tree Can be used as regular array, but with two imp

Build a Segment Tree data structure. https://en.wikipedia.org/wiki/Segment_tree Can be used as regular array, but with two imp

Represents world in a `height x width x 6` `Tensor`.

Clean up the environments' resources.

Mirror an observation, not including the time and key bars.

Run training indefinitely, saving periodically.

Create a rollout of segments from recordings. Args: recordings: a sequence of recordings. batch: the number of segments to g