src.agents.petsplanners package

Submodules

src.agents.petsplanners.CEMPlanner module

CEM Planner for PETS. Note that we’ve had some errors with this in the past, due to not tuning the parameters properly.

class src.agents.petsplanners.CEMPlanner.CEMPlanner(action_dim: int = 2, action_min: float = -1, action_max: float = 1, n_planner: int = 500, horizon: int = 12, iter_update_steps: int = 3, k_best: int = 5, epsilon: float = 0.001, update_alpha: float = 0.1, lb: float = -0.5, ub: float = 0.5)

Bases: BasePlanner

CEM ( Gaussian Evolutionary Method ) Based Planner. Untuned.

Initialize CEM Planner

Parameters

  • action_dim (int, optional) – Number of dimensions in action space. Defaults to 2.

  • action_min (float, optional) – Minimum action value. Defaults to -1.

  • action_max (float, optional) – Maximum action value. Defaults to 1.

  • n_planner (int, optional) – Number of trajectories to plan. Defaults to 500.

  • horizon (int, optional) – How many steps in the future to plan. Defaults to 12.

  • iter_update_steps (int, optional) – How many times to iterate on the planning side. Defaults to 3.

  • k_best (int, optional) – Number of best trajectories to keep. Defaults to 5.

  • epsilon (float, optional) – Minimum variance of planner. Defaults to 1e-3.

  • update_alpha (float, optional) – How much to update planner based on results. Defaults to 0.1.

  • lb (float, optional) – Lower bound. Defaults to -0.5.

  • ub (float, optional) – Upper bound. Defaults to 0.5.

get_action(state, dynamics_model: Module) → array

Generate action given state and dynamics model

Parameters

  • state (torch.Tensor) – State to plan from

  • dynamics_model (torch.nn.Module) – Dynamics model ( state, action -> nextstate, rewards )

  • noise (bool, optional) – Whether to add noise to action. Defaults to False.

Returns

Planned action

Return type

np.array

classmethod instantiate_from_config(config_file_location)

Initialize class from config file

Parameters

config_file_location (path) – Path to config file

Raises

ValueError – Error loading file

Returns

object from class.

Return type

cls

classmethod instantiate_from_config_dict(config)

Initialize class from config dictionary

Parameters

config (dictionary) – Create instance of class from dictionary.

Returns

Object from class and config.

Return type

cls

schema = Map({Optional("action_dim"): Int(), Optional("action_min"): Float(), Optional("action_max"): Float(), Optional("n_planner"): Int(), Optional("horizon"): Int(), Optional("iter_update_steps"): Int(), Optional("k_best"): Int(), Optional("epsilon"): Float(), Optional("update_alpha"): Float(), Optional("lb"): Float(), Optional("ub"): Float()})

src.agents.petsplanners.PDDMPlanner module

PDDM Planner. Largely untested.

class src.agents.petsplanners.PDDMPlanner.PDDMPlanner(action_dim: int = 2, action_min: float = -1, action_max: float = 1, n_planner: int = 500, horizon: int = 12, gamma: float = 1.0, beta: float = 0.5)

Bases: BasePlanner

Likely an iteration of https://github.com/google-research/pddm.

Initialize PDDM Planner

Parameters

  • action_dim (int, optional) – Size of action space. Defaults to 2.

  • action_min (float, optional) – Minimum action value. Defaults to -1.

  • action_max (float, optional) – Maximum action value. Defaults to 1.

  • n_planner (int, optional) – Number of trajectories to plan. Defaults to 500.

  • horizon (int, optional) – Length to plan into the future. Defaults to 12.

  • gamma (float, optional) – Gamma. Defaults to 1.0.

  • beta (float, optional) – Beta. Defaults to 0.5.

get_action(state, dynamics_model: Module, noise: bool = True) → array

Generate action given state and dynamics model

Parameters

  • state (torch.Tensor) – State to plan from

  • dynamics_model (torch.nn.Module) – Dynamics model ( state, action -> nextstate, rewards )

  • noise (bool, optional) – Whether to add noise to action. Defaults to False.

Returns

Planned action

Return type

np.array

classmethod instantiate_from_config(config_file_location)

Initialize class from config file

Parameters

config_file_location (path) – Path to config file

Raises

ValueError – Error loading file

Returns

object from class.

Return type

cls

classmethod instantiate_from_config_dict(config)

Initialize class from config dictionary

Parameters

config (dictionary) – Create instance of class from dictionary.

Returns

Object from class and config.

Return type

cls

schema = Map({Optional("action_dim"): Int(), Optional("action_min"): Float(), Optional("action_max"): Float(), Optional("n_planner"): Int(), Optional("horizon"): Int(), Optional("gamma"): Float(), Optional("beta"): Float()})

src.agents.petsplanners.RandomPlanner module

PDDM Planner. Largely untested.

class src.agents.petsplanners.RandomPlanner.RandomPlanner(action_dim: int = 2, action_min: float = -1, action_max: float = 1, n_planner: int = 500, horizon: int = 12)

Bases: BasePlanner

Random Shooting-Based Planner.

Initialize Random Shooting Planner

Parameters

  • action_dim (int, optional) – Dimension of action space. Defaults to 2.

  • action_min (float, optional) – Minimum action. Defaults to -1.

  • action_max (float, optional) – Maximum action. Defaults to 1.

  • n_planner (int, optional) – Number of trajectories to test. Defaults to 500.

  • horizon (int, optional) – Planning horizon. Defaults to 12.

get_action(state, dynamics_model: Module, noise: bool = False) → array

Generate action given state and dynamics model

Parameters

  • state (torch.Tensor) – State to plan from

  • dynamics_model (torch.nn.Module) – Dynamics model ( state, action -> nextstate, rewards )

  • noise (bool, optional) – Whether to add noise to action. Defaults to False.

Returns

Planned action

Return type

np.array

classmethod instantiate_from_config(config_file_location)

Initialize class from config file

Parameters

config_file_location (path) – Path to config file

Raises

ValueError – Error loading file

Returns

object from class.

Return type

cls

classmethod instantiate_from_config_dict(config)

Initialize class from config dictionary

Parameters

config (dictionary) – Create instance of class from dictionary.

Returns

Object from class and config.

Return type

cls

schema = Map({Optional("action_dim"): Int(), Optional("action_min"): Float(), Optional("action_max"): Float(), Optional("n_planner"): Int(), Optional("horizon"): Int()})

src.agents.petsplanners.base module

class src.agents.petsplanners.base.BasePlanner(n_planner: int = 500, horizon: int = 12)

Bases: ABC

get_action(state) → array

Generate action given state and dynamics model

Parameters

  • state (torch.Tensor) – State to plan from

  • dynamics_model (torch.nn.Module) – Dynamics model ( state, action -> nextstate, rewards )

  • noise (bool, optional) – Whether to add noise to action. Defaults to False.

Returns

Planned action

Return type

np.array

Module contents