def softmax_selection(self, state_index, Q):
possible_actions = self.env.get_possible_actions(state_index)
if len(possible_actions) == 1:
return 0
probabilities = utils.softmax(Q, self.inv_temp)
action_idx = np.random.choice(list(range(len(possible_actions))), p=probabilities)
return action_idx
def select_action(self, state_idx, softmax=True):
# TODO: get categorical dist over next state
# okay because it's local
# gradient-based (hill-climbing) gradient ascent
# graph hill climbing
# Maybe change for M(sa,sa). potentially over state action only in two step
V = self.M_hat @ self.R_hat
next_state = [self.env.get_next_state(state_idx, a) for a in range(self.env.nr_actions)]
Q = [V[s] for s in next_state]
probabilities = utils.softmax(Q, self.beta)
return np.random.choice(list(range(self.env.nr_actions)), p=probabilities)
def select_action(self, state_idx, softmax=True):
# TODO: get categorical dist over next state
# okay because it's local
# gradient-based planning (hill-climbing) gradient ascent
# graph hill climbing
# Maybe change for M(sa,sa). potentially over state action only in two step
next_state = [
self.env.get_next_state(state_idx, a)
for a in range(self.env.nr_actions)
]
Q = [self.compute_V(s) for s in next_state]
probabilities = utils.softmax(Q, self.beta)
try:
a = np.random.choice(list(range(self.env.nr_actions)),
p=probabilities)
except ValueError:
print('whats wrong')
return a
def softmax_selection(self, state_index, Q):
probabilities = utils.softmax(Q, self.inv_temp)
action_idx = np.random.choice(list(range(self.env.nr_actions)),
p=probabilities)
return action_idx
def softmax_selection(self, state_idx):
probabilities = utils.softmax(self.Q[state_idx], self.beta)
action_idx = np.random.choice(list(range(self.env.nr_actions)),
p=probabilities)
return action_idx
