def get_reward(self, _state, _action, _next_state):
"""
Gets the reward for the state, action, next_state transition.
Note that this is not available in reinforcement learning.
"""
abstract()
def get_transition_states_and_probs(self, _state, _action):
"""
Returns a list of (next_state, prob) pairs representing the states
reachable from 'state' by taking 'action' along with their transition
probabilities.
Note that in Q-Learning and reinforcement learning in general, we
won't know these probabilities nor do we directly model them.
"""
abstract()
def is_terminal(self, _state):
"""
Returns true if the current state is a terminal state.
Note that, by convention, a terminal state has zero future rewards.
It's possible that the terminal state(s) have no possible actions.
It's also common to think of the terminal state as having a type of
self-loop action 'pass' with zero reward. Those formulations are
essentially equivalent.
"""
abstract()
def get_start_state(self):
"""
Provides the starting state of the MDP.
"""
abstract()
def get_possible_actions(self, _state):
"""
Provides a list of possible actions from the provided state.
"""
abstract()
def get_states(self):
"""
Provides a list of all states in the MDP. Note that for very
large MDPs this may not possible.
"""
abstract()
def do_action(self, _action):
"""
Performs the given action in the current environment state and
updates the environment. Returns a (reward, next_state) pair.
"""
abstract()
def reset(self):
"""
Returns the environment to its start state.
"""
abstract()
def get_possible_actions(self, _state):
"""
Returns the possible actions an agent can take in the given state.
An empty list is returned if the environment is in a terminal state.
"""
abstract()
def get_current_state(self):
"""
Returns the current state of the environment.
"""
abstract()