def reset(self,
uniform=False
): # initializes an episode and returns the state of the agent
# if uniform is set to False, the first state is drawn according to the P0 distribution,
# else it is drawn from a uniform distribution over all the states except for walls
if uniform:
prob = np.ones(self.nb_states) / self.nb_states
self.current_state = discreteProb(prob)
else:
self.current_state = discreteProb(self.P0)
self.timestep = 0
self.last_action_achieved = False
return self.current_state
def step(self,
u,
deviation=0): # performs a step forward in the environment,
# if you want to add some noise to the reward, give a value to the deviation param
# which represents the mean μ of the normal distribution used to draw the noise
noise = 0 # = deviation*np.random.randn() # generate noise, see an exercize in mbrl.ipynb
reward = self.r[
self.current_state,
u] + noise # r is the reward of the transition, you can add some noise to it
# the state reached when performing action u from state x is sampled
# according to the discrete distribution self.P[x,u,:]
observation = discreteProb(self.P[self.current_state, u, :])
self.timestep += 1
info = {} #can be used when debugging
info["State transition probabilities"] = self.P[self.current_state,
u, :]
info["reward's noise value"] = noise
self.current_state = observation
done = self.done() #checks if the episode is over
return [observation, reward, done, info]
def reset(self, uniform=False): # initializes an episode
# if uniform is set to False, the first state is drawn from the P0 distribution,
# else it is drawn from a uniform distribution over all the states except for walls
if uniform:
prob = np.ones(self.observation_space.size) / (
self.observation_space.size -
len(self.observation_space.walls))
for state in self.observation_space.walls:
prob[state] = 0.0
self.current_state = discreteProb(prob)
else:
self.current_state = discreteProb(self.P0)
self.timestep = 0
self.last_action_achieved = False
return self.current_state
def step(self,
u,
deviation=0): # performs a step forward in the environment,
# if you want to add some noise to the reward, give a value to the deviation param
# which represents the mean μ of the normal distribution used to draw the noise
noise = deviation * np.random.randn(
) # generate noise, useful for RTDP
# r is the reward of the transition, you can add some noise to it
reward = self.r[self.current_state, u] + noise
# the state reached when performing action u from state x is sampled
# according to the discrete distribution self.P[x,u,:]
state = discreteProb(self.P[self.current_state, u, :])
self.timestep += 1
info = {
"State transition probabilities": self.P[self.current_state, u, :],
"reward's noise value": noise
} # can be used when debugging
self.current_state = state
done = self.done() # checks if the episode is over
return [state, reward, done, info]
def sample(self, prob_list=None):
# returns an action drawn according to the prob_list distribution,
# if the param is not set, then it is drawn from a uniform distribution
if prob_list is None:
prob_list = np.ones(self.size) / self.size
index = discreteProb(prob_list)
return self.actions[index]
