MuJoCo problem

class blackbox.mujoco_problems.MujocoProblem(env_name, lb=-1.0, ub=1.0, num_rollouts=3, seed=None)[source]

Bases: BaseProblem

__call__(x)[source]

Method to compute total reward for a given policy x

Parameters:

x (np.ndarray) – input to be evaluated, can be a single sample or multiple samples, (dim,) or (n_samples,dim)

Returns:

reward – reward of the trajectory for the given input, (1,) or (n_samples,1)

Return type:

np.ndarray

__init__(env_name, lb=-1.0, ub=1.0, num_rollouts=3, seed=None)[source]

Class for setting up MuJoCo environment and run the simulation

Parameters:
  • env_name (str) – name of the mujoco environment, must be: ‘Swimmer-v5’, ‘HalfCheetah-v5’, ‘Hopper-v5’, ‘Walker2d-v5’, or ‘Humanoid-v5’

  • lb (float, default=-1.0) – lower bound of the input space

  • ub (float, default=1.0) – upper bound of the input space

  • num_rollouts (int, default=3) – number of times to run the entire episode for a given policy

  • seed (int | None, default=None) – seed for the environment

class blackbox.mujoco_problems.RunningStat(shape=None)[source]

Bases: object

__init__(shape=None)[source]

Class for computing statistics for a stream of data, here it is used for computing the mean and standard deviation of the observation space of the Mujoco env. This mean and standard deviation is used for normalizing the observation space before computing the next action

http://www.johndcook.com/blog/standard_deviation/

property mean
property n
push(x)[source]

Method for adding a new value of x i.e new observation state to the running statistics

Note: x should be a single observation state, not a batch of states

property shape
property std
property var