Core Types

ActionInfo

class rl_coach.core_types.ActionInfo(action: Union[int, float, numpy.ndarray, List], all_action_probabilities: float = 0, action_value: float = 0.0, state_value: float = 0.0, max_action_value: float = None)

Action info is a class that holds an action and various additional information details about it

Parameters

• action – the action

• all_action_probabilities – the probability that the action was given when selecting it

• action_value – the state-action value (Q value) of the action

• state_value – the state value (V value) of the state where the action was taken

• max_action_value – in case this is an action that was selected randomly, this is the value of the action that received the maximum value. if no value is given, the action is assumed to be the action with the maximum value

Batch

class rl_coach.core_types.Batch(transitions: List[rl_coach.core_types.Transition])

A wrapper around a list of transitions that helps extracting batches of parameters from it. For example, one can extract a list of states corresponding to the list of transitions. The class uses lazy evaluation in order to return each of the available parameters.

Parameters

transitions – a list of transitions to extract the batch from

actions(expand_dims=False) → numpy.ndarray

if the actions were not converted to a batch before, extract them to a batch and then return the batch

Parameters

expand_dims – add an extra dimension to the actions batch

Returns

a numpy array containing all the actions of the batch

game_overs(expand_dims=False) → numpy.ndarray

if the game_overs were not converted to a batch before, extract them to a batch and then return the batch

Parameters

expand_dims – add an extra dimension to the game_overs batch

Returns

a numpy array containing all the game over flags of the batch

goals(expand_dims=False) → numpy.ndarray

if the goals were not converted to a batch before, extract them to a batch and then return the batch if the goal was not filled, this will raise an exception

Parameters

expand_dims – add an extra dimension to the goals batch

Returns

a numpy array containing all the goals of the batch

info(key, expand_dims=False) → numpy.ndarray

if the given info dictionary key was not converted to a batch before, extract it to a batch and then return the batch. if the key is not part of the keys in the info dictionary, this will raise an exception

Parameters

expand_dims – add an extra dimension to the info batch

Returns

a numpy array containing all the info values of the batch corresponding to the given key

info_as_list(key) → list

get the info and store it internally as a list, if wasn’t stored before. return it as a list :param expand_dims: add an extra dimension to the info batch :return: a list containing all the info values of the batch corresponding to the given key

n_step_discounted_rewards(expand_dims=False) → numpy.ndarray

if the n_step_discounted_rewards were not converted to a batch before, extract them to a batch and then return

the batch

if the n step discounted rewards were not filled, this will raise an exception :param expand_dims: add an extra dimension to the total_returns batch :return: a numpy array containing all the total return values of the batch

next_states(fetches: List[str], expand_dims=False) → Dict[str, numpy.ndarray]

follow the keys in fetches to extract the corresponding items from the next states in the batch if these keys were not already extracted before. return only the values corresponding to those keys

Parameters

• fetches – the keys of the state dictionary to extract

• expand_dims – add an extra dimension to each of the value batches

Returns

a dictionary containing a batch of values correponding to each of the given fetches keys

rewards(expand_dims=False) → numpy.ndarray

if the rewards were not converted to a batch before, extract them to a batch and then return the batch

Parameters

expand_dims – add an extra dimension to the rewards batch

Returns

a numpy array containing all the rewards of the batch

shuffle() → None

Shuffle all the transitions in the batch

Returns

None

property size

Returns

the size of the batch

slice(start, end) → None

Keep a slice from the batch and discard the rest of the batch

Parameters

• start – the start index in the slice

• end – the end index in the slice

Returns

None

states(fetches: List[str], expand_dims=False) → Dict[str, numpy.ndarray]

follow the keys in fetches to extract the corresponding items from the states in the batch if these keys were not already extracted before. return only the values corresponding to those keys

Parameters

• fetches – the keys of the state dictionary to extract

• expand_dims – add an extra dimension to each of the value batches

Returns

a dictionary containing a batch of values correponding to each of the given fetches keys

EnvResponse

class rl_coach.core_types.EnvResponse(next_state: Dict[str, numpy.ndarray], reward: Union[int, float, numpy.ndarray], game_over: bool, info: Dict = None, goal: numpy.ndarray = None)

An env response is a collection containing the information returning from the environment after a single action has been performed on it.

Parameters

• next_state – The new state that the environment has transitioned into. Assumed to be a dictionary where the observation is located at state[‘observation’]

• reward – The reward received from the environment

• game_over – A boolean which should be True if the episode terminated after the execution of the action.

• info – any additional info from the environment

• goal – a goal defined by the environment

Episode

class rl_coach.core_types.Episode(discount: float = 0.99, bootstrap_total_return_from_old_policy: bool = False, n_step: int = -1)

An Episode represents a set of sequential transitions, that end with a terminal state.

Parameters

• discount – the discount factor to use when calculating total returns

• bootstrap_total_return_from_old_policy – should the total return be bootstrapped from the values in the memory

• n_step – the number of future steps to sum the reward over before bootstrapping

get_first_transition() → rl_coach.core_types.Transition

Get the first transition in the episode, or None if there are no transitions available

Returns

The first transition in the episode

get_last_transition() → rl_coach.core_types.Transition

Get the last transition in the episode, or None if there are no transition available

Returns

The last transition in the episode

get_transition(transition_idx: int) → rl_coach.core_types.Transition

Get a specific transition by its index.

Parameters

transition_idx – The index of the transition to get

Returns

The transition which is stored in the given index

get_transitions_attribute(attribute_name: str) → List[Any]

Get the values for some transition attribute from all the transitions in the episode. For example, this allows getting the rewards for all the transitions as a list by calling get_transitions_attribute(‘reward’)

Parameters

attribute_name – The name of the attribute to extract from all the transitions

Returns

A list of values from all the transitions according to the attribute given in attribute_name

insert(transition: rl_coach.core_types.Transition) → None

Insert a new transition to the episode. If the game_over flag in the transition is set to True, the episode will be marked as complete.

Parameters

transition – The new transition to insert to the episode

Returns

None

is_empty() → bool

Check if the episode is empty

Returns

A boolean value determining if the episode is empty or not

length() → int

Return the length of the episode, which is the number of transitions it holds.

Returns

The number of transitions in the episode

update_discounted_rewards()

Update the discounted returns for all the transitions in the episode. The returns will be calculated according to the rewards of each transition, together with the number of steps to bootstrap from and the discount factor, as defined by n_step and discount respectively when initializing the episode.

Returns

None

Transition

class rl_coach.core_types.Transition(state: Dict[str, numpy.ndarray] = None, action: Union[int, float, numpy.ndarray, List] = None, reward: Union[int, float, numpy.ndarray] = None, next_state: Dict[str, numpy.ndarray] = None, game_over: bool = None, info: Dict = None)

A transition is a tuple containing the information of a single step of interaction between the agent and the environment. The most basic version should contain the following values: (current state, action, reward, next state, game over) For imitation learning algorithms, if the reward, next state or game over is not known, it is sufficient to store the current state and action taken by the expert.

Parameters

• state – The current state. Assumed to be a dictionary where the observation is located at state[‘observation’]

• action – The current action that was taken

• reward – The reward received from the environment

• next_state – The next state of the environment after applying the action. The next state should be similar to the state in its structure.

• game_over – A boolean which should be True if the episode terminated after the execution of the action.

• info – A dictionary containing any additional information to be stored in the transition