colosseum.mdp.simple_grid.finite_horizon

View Source

 1from typing import Any, Dict, List
 2
 3import gin
 4
 5from colosseum.mdp import EpisodicMDP
 6from colosseum.mdp.simple_grid.base import SimpleGridMDP
 7
 8
 9@gin.configurable
10class SimpleGridEpisodic(EpisodicMDP, SimpleGridMDP):
11    @staticmethod
12    def sample_parameters(n: int, seed: int = None) -> List[Dict[str, Any]]:
13        return SimpleGridMDP.sample_mdp_parameters(n, True, seed)

@gin.configurable

class SimpleGridEpisodic(colosseum.mdp.base_finite.EpisodicMDP, colosseum.mdp.simple_grid.base.SimpleGridMDP): View Source

10@gin.configurable
11class SimpleGridEpisodic(EpisodicMDP, SimpleGridMDP):
12    @staticmethod
13    def sample_parameters(n: int, seed: int = None) -> List[Dict[str, Any]]:
14        return SimpleGridMDP.sample_mdp_parameters(n, True, seed)

The base class for episodic MDPs.

SimpleGridEpisodic(H: int = None, **kwargs) View Source

 78    def __init__(self, H: int = None, **kwargs):
 79        super(EpisodicMDP, self).__init__(**kwargs)
 80
 81        # Computing the time horizon
 82        self._input_H = H
 83        self._H = None
 84
 85        # Episodic setting specific caching variables
 86        self._reachable_states = None
 87        self._episodic_graph = dict()
 88        self._continuous_form_episodic_transition_matrix_and_rewards = None
 89        self._episodic_transition_matrix_and_rewards = None
 90        self._optimal_policy_cf = dict()
 91        self._worst_policy_cf = dict()
 92        self._optimal_value_cf = None
 93        self._worst_value_cf = None
 94        self._random_value_cf = None
 95        self._eoar = None
 96        self._woar = None
 97        self._roar = None
 98        self._random_policy_cf = None
 99        self._random_policy = None
100        self._average_optimal_episodic_reward = None
101        self._average_worst_episodic_reward = None
102        self._average_random_episodic_reward = None

Parameters

seed (int): The seed used for sampling rewards and next states.
size (int): The size of the grid.
reward_type (SimpleGridReward): The type of reward for the MDP. By default, the XOR type is used.
n_starting_states (int): The number of possible starting states.
optimal_mean_reward (float): If the rewards are made stochastic, this parameter controls the mean reward for the optimal trajectory. By default, it is set to 0.9.
sub_optimal_mean_reward (float): If the rewards are made stochastic, this parameter controls the mean reward for suboptimal trajectories. By default, it is set to 0.2.
optimal_distribution (Union[Tuple, rv_continuous]): The distribution of the highly rewarding state. It can be either passed as a tuple containing Beta parameters or as a rv_continuous object.
sub_optimal_distribution (Union[Tuple, rv_continuous]): The distribution of the suboptimal rewarding states. It can be either passed as a tuple containing Beta parameters or as a rv_continuous object.
other_distribution (Union[Tuple, rv_continuous]): The distribution of the other states. It can be either passed as a tuple containing Beta parameters or as a rv_continuous object.
make_reward_stochastic (bool): If True, the rewards of the MDP will be stochastic. By default, it is set to False.
reward_variance_multiplier (float): A constant that can be used to increase the variance of the reward distributions without changing their means. The lower the value, the higher the variance. By default, it is set to 1.

@staticmethod

def sample_parameters(n: int, seed: int = None) -> List[Dict[str, Any]]: View Source

12    @staticmethod
13    def sample_parameters(n: int, seed: int = None) -> List[Dict[str, Any]]:
14        return SimpleGridMDP.sample_mdp_parameters(n, True, seed)

Returns

List[Dict[str, Any]]: n sampled parameters that can be used to construct an MDP in a reasonable amount of time.

Inherited Members

colosseum.mdp.base_finite.EpisodicMDP: is_episodic; H; random_policy_cf; random_policy; parameters; reachable_states; T_cf; R_cf; optimal_value_continuous_form; worst_value_continuous_form; random_value_continuous_form; episodic_optimal_average_reward; episodic_worst_average_reward; episodic_random_average_reward; continuous_form_episodic_transition_matrix_and_rewards; episodic_transition_matrix_and_rewards; get_optimal_policy_continuous_form; get_worst_policy_continuous_form; get_random_policy_continuous_form; get_minimal_regret_for_starting_node; get_optimal_policy_starting_value; get_worst_policy_starting_value; get_random_policy_starting_value; get_episodic_graph; get_grid_representation
colosseum.mdp.simple_grid.base.SimpleGridMDP: get_action_class; get_unique_symbols; does_seed_change_MDP_structure; sample_mdp_parameters; get_node_class; get_gin_parameters; n_actions
colosseum.mdp.base.BaseMDP: get_available_hardness_measures; produce_gin_file_from_mdp_parameters; get_gin_config; get_node_labels; get_node_action_labels; hash; instantiate_MDP; T; R; recurrent_nodes_set; communication_class; get_optimal_policy; get_worst_policy; get_value_functions; optimal_value_functions; worst_value_functions; random_value_functions; optimal_transition_probabilities; worst_transition_probabilities; random_transition_probabilities; optimal_markov_chain; worst_markov_chain; random_markov_chain; get_stationary_distribution; optimal_stationary_distribution; worst_stationary_distribution; random_stationary_distribution; optimal_average_rewards; worst_average_rewards; random_average_rewards; get_average_reward; optimal_average_reward; worst_average_reward; random_average_reward; transition_matrix_and_rewards; graph_layout; graph_metrics; diameter; sum_reciprocals_suboptimality_gaps; discounted_value_norm; undiscounted_value_norm; value_norm; measures_of_hardness; summary; hardness_report; get_info_class; get_transition_distributions; get_reward_distribution; sample_reward; get_measure_from_name; action_spec; observation_spec; get_observation; reset; step; random_steps; random_step; get_visitation_counts; reset_visitation_counts; get_value_node_labels
dm_env._environment.Environment: reward_spec; discount_spec; close