class SARSALambdaDiscrete(TD):
"""
Discrete version of SARSA(lambda) algorithm.
"""
def __init__(self, policy, mdp_info, params):
self.Q = Table(mdp_info.size)
self._lambda = params['algorithm_params']['lambda']
trace = params.get('algorithm_params', 'replacing')
self.e = EligibilityTrace(self.Q.shape, trace)
super(SARSALambdaDiscrete, self).__init__(self.Q, policy, mdp_info,
params)
def _update(self, state, action, reward, next_state, absorbing):
q_current = self.Q[state, action]
self._next_action = self.draw_action(next_state)
q_next = self.Q[next_state, self._next_action] if not absorbing else 0.
delta = reward + self.mdp_info.gamma * q_next - q_current
self.e.update(state, action)
self.Q.table += self.alpha(state, action) * delta * self.e.table
self.e.table *= self.mdp_info.gamma * self._lambda
def episode_start(self):
self.e.reset()
def __init__(self, policy, mdp_info, params):
self.Q = Table(mdp_info.size)
self._lambda = params['algorithm_params']['lambda']
trace = params.get('algorithm_params', 'replacing')
self.e = EligibilityTrace(self.Q.shape, trace)
super(SARSALambdaDiscrete, self).__init__(self.Q, policy, mdp_info,
params)
def __init__(self,
policy,
mdp_info,
learning_rate,
lambda_coeff,
trace='replacing'):
"""
Constructor.
Args:
lambda_coeff (float): eligibility trace coefficient;
trace (str, 'replacing'): type of eligibility trace to use.
"""
self.Q = Table(mdp_info.size)
self._lambda = lambda_coeff
self.e = EligibilityTrace(self.Q.shape, trace)
super().__init__(self.Q, policy, mdp_info, learning_rate)
class SARSALambdaDiscrete(TD):
"""
Discrete version of SARSA(lambda) algorithm.
"""
def __init__(self,
policy,
mdp_info,
learning_rate,
lambda_coeff,
trace='replacing'):
"""
Constructor.
Args:
lambda_coeff (float): eligibility trace coefficient;
trace (str, 'replacing'): type of eligibility trace to use.
"""
self.Q = Table(mdp_info.size)
self._lambda = lambda_coeff
self.e = EligibilityTrace(self.Q.shape, trace)
super().__init__(self.Q, policy, mdp_info, learning_rate)
def _update(self, state, action, reward, next_state, absorbing):
q_current = self.Q[state, action]
self.next_action = self.draw_action(next_state)
q_next = self.Q[next_state, self.next_action] if not absorbing else 0.
delta = reward + self.mdp_info.gamma * q_next - q_current
self.e.update(state, action)
self.Q.table += self.alpha(state, action) * delta * self.e.table
self.e.table *= self.mdp_info.gamma * self._lambda
def episode_start(self):
self.e.reset()
super().episode_start()
class SARSALambdaDiscrete(TD):
"""
Discrete version of SARSA(lambda) algorithm.
"""
def __init__(self, policy, mdp_info, learning_rate, lambda_coeff,
trace='replacing'):
"""
Constructor.
Args:
lambda_coeff (float): eligibility trace coefficient;
trace (str, 'replacing'): type of eligibility trace to use.
"""
self.Q = Table(mdp_info.size)
self._lambda = lambda_coeff
self.e = EligibilityTrace(self.Q.shape, trace)
super(SARSALambdaDiscrete, self).__init__(self.Q, policy, mdp_info,
learning_rate)
def _update(self, state, action, reward, next_state, absorbing):
q_current = self.Q[state, action]
self.next_action = self.draw_action(next_state)
q_next = self.Q[next_state, self.next_action] if not absorbing else 0.
delta = reward + self.mdp_info.gamma * q_next - q_current
self.e.update(state, action)
self.Q.table += self.alpha(state, action) * delta * self.e.table
self.e.table *= self.mdp_info.gamma * self._lambda
def episode_start(self):
self.e.reset()
def __init__(self, policy, mdp_info, learning_rate, lambda_coeff,
trace='replacing'):
"""
Constructor.
Args:
lambda_coeff (float): eligibility trace coefficient;
trace (str, 'replacing'): type of eligibility trace to use.
"""
self.Q = Table(mdp_info.size)
self._lambda = lambda_coeff
self.e = EligibilityTrace(self.Q.shape, trace)
super(SARSALambdaDiscrete, self).__init__(self.Q, policy, mdp_info,
learning_rate)