Compute the observations of all environment. The observation is composed of three parts: the state values of the left and right hands

Fisher vector product.

Flatten the gradients.

Flatten the parameters.

Retrieves param value for the given prop and attr from the sample.

Returns the available actions of all agents in a list.

Get action logits from inputs. Args: x: (torch.Tensor) input to network. available_actions: (torch.Tensor) denotes whi

Get the number of agents in the environment.

Returns all agent observations in a list.

Returns observation for agent_id. The observation is composed of: - agent movement features (where it can move to, height information and pat

Returns the dimensions of the matrix containing ally features. Size is n_allies x n_features.

Returns the dimensions of the matrix containing enemy features. Size is n_enemies x n_features.

Returns the size of the vector containing the agents' own features.

Returns the shape of the observation

Aggregate reward functions Args: env: environment instance agent_id: current agent id info: addi

Get shape from action space. Args: act_space: (gym.spaces) action space Returns: act_shape: (tuple) action shape

Returns the global state.

Returns the global state. NOTE: This functon should not be used during decentralised execution.

Returns the dimensions of the matrix containing ally features. Size is n_allies x n_features.

Returns the dimensions of the matrix containing enemy features. Size is n_enemies x n_features.

Returns the size of the vector containing the agents's movement-related features.

Returns the size of the vector containing the agents's movement-related features. global

Returns the size of the vector containing the agents' own features.

Returns the size of the global state.

Returns the shape of the state

Returns the surrounding points of the unit in 8 directions.

Get target actor actions for observations. Args: obs: (np.ndarray) observations of target actor, shape is (batch_size, dim)

Get task name.

Aggregate termination conditions Args: env: environment instance agent_id: current agent id info

Sample from the Gumbel-Softmax distribution and optionally discretize. Args: logits: [batch_size, n_class] unnormalized log-probs temp

Huber loss.

Given an angle in rads, normalises in (-pi, pi]

Init device. Args: args: (dict) arguments Returns: device: (torch.device) device

Init directory for saving results.

KL divergence between two distributions.

Make env for evaluation.

Make env for rendering.

Make env for training.

MSE loss.

Normalize action to be consistent with action space.

Check if only Medivac units are left.

Prepare for rollout.

Get running mean and variance.

Save the configuration of the program.

Set the values of the specified properties :param props: list of Properties :param values: list of float

Seed the program.

Smooth signal y, where radius is determines the size of the window. mode='twosided': average over the window [max(index - radius, 0), min

Step the environments synchronously. This is available for backwards compatibility.

A single environment step. Returns reward, terminated, info.

Get values with grad for obs and actions Args: obs: (np.ndarray) observations batch, shape is (batch_size, dim) action

Update actor network. Args: sample: (Tuple) contains data batch with which to update networks. Returns: policy