def agent_step(
reward, this_observation
): # returns NumPy array, reward: floating point, observation_t: NumPy array
global local_action, last_observation, this_action, action_value_estimates, action_counts, time_step, C
#Update estimate for current action
cur_action = int(this_action[0])
action_value_estimates[cur_action] += alpha * (
reward - action_value_estimates[cur_action])
#Choose a new action using the parameter based agents
stp1 = this_observation[0]
action_selection_prob = rand_un()
if action_selection_prob <= (1 - epsilon):
atp1 = action_value_estimates.index(max(action_value_estimates))
elif episode == EPSILON_GREEDY:
atp1 = randInRange(numActions)
else:
print("BAD EPISODE: NO ACTION SELECTION FOR THE CURRENT AGENT!!!")
exit()
action_counts[atp1] += 1
time_step += 1
local_action[0] = atp1
this_action = local_action
last_observation = this_observation
return this_action
def agent_start(this_observation
): # returns NumPy array, this_observation: NumPy array
global local_action, last_observation, this_action, episode, action_value_estimates, action_counts, epsilon, Q1, time_step
#Set the parameters based on the current agent
#We use episodes to distinguish between agent parameter settings
if episode == OPTIMISTIC_INIT:
epsilon = 0.0
Q1 = 5
elif episode == EPSILON_GREEDY:
epsilon = 0.1
Q1 = 0
else:
exit("BAD EPISODE: NO STRATEGY FOR THE CURRENT AGENT!!!")
action_value_estimates = [Q1 for action in range(numActions)]
action_counts = [0 for action in range(numActions)]
stp1 = this_observation[
0] # how you convert observation to a number, if state is tabular
atp1 = randInRange(numActions)
action_counts[atp1] += 1
local_action[0] = atp1
last_observation = this_observation # save observation, might be useful on the next step
this_action = local_action
time_step = 1
return this_action
def agent_start(this_observation): # returns NumPy array, this_observation: NumPy array
global local_action, last_observation, this_action#, numActions
stp1 = this_observation[0] # how you convert observation to a number, if state is tabular
atp1 = randInRange(numActions)
local_action[0] = atp1
last_observation = this_observation # save observation, might be useful on the next step
this_action = local_action
return this_action
def agent_step(reward, this_observation): # returns NumPy array, reward: floating point, observation_t: NumPy array
global local_action, last_observation, this_action#, numActions
stp1 = this_observation[0]
atp1 = randInRange(numActions)
# might do some learning here
local_action[0] = atp1
this_action = local_action
last_observation = this_observation
return this_action
def env_step(this_action): # returns (floating point, NumPy array, Boolean), this_action: NumPy array
global local_observation, this_reward_observation, arms#, nStatesSimpleEnv
episode_over = False
atp1 = this_action[0] # how to extact action
stp1 = randInRange(nStatesSimpleEnv) # state transitions are uniform random
the_reward = randn(0.0, 1.0) + arms[int(atp1)] # rewards drawn from (0, 1) Gaussian
#if rand_un() < 0.05:
# episode_over = True # termination is random
local_observation[0] = stp1
this_reward_observation = (the_reward, this_reward_observation[1], episode_over)
return this_reward_observation
def agent_step(
reward, this_observation
): # returns NumPy array, reward: floating point, observation_t: NumPy array
global local_action, last_observation, this_action, action_value_estimates, action_counts, time_step, C
#Update estimate for current action
cur_action = int(this_action[0])
action_value_estimates[cur_action] += alpha * (
reward - action_value_estimates[cur_action])
#Choose a new action using the parameter based agents
stp1 = this_observation[0]
action_selection_prob = rand_un()
if episode == EPSILON_GREEDY:
if action_selection_prob <= (1 - epsilon):
atp1 = action_value_estimates.index(max(action_value_estimates))
else:
atp1 = randInRange(numActions)
elif episode == UCB:
action_value_estimates_copy = list(action_value_estimates)
cur_greedy_action_index = action_value_estimates_copy.index(
max(action_value_estimates_copy))
action_counts_copy = list(action_counts)
#Compute the uncertainty measure for each action value estimate, and select the next action
action_values_UCB = []
for i in range(len(action_value_estimates_copy)):
cur_UCB_action_value = action_value_estimates_copy[i] + (
C *
(math.sqrt(math.log(time_step) / (action_counts_copy[i] + 1))))
action_values_UCB.append(cur_UCB_action_value)
atp1 = action_values_UCB.index(max(action_values_UCB))
else:
exit("BAD EPISODE: NO ACTION SELECTION FOR THE CURRENT AGENT!!!")
action_counts[atp1] += 1
time_step += 1
local_action[0] = atp1
this_action = local_action
last_observation = this_observation
return this_action
def agent_start(this_observation
): # returns NumPy array, this_observation: NumPy array
global local_action, last_observation, this_action, episode, action_value_estimates, action_counts, epsilon, Q1, time_step
action_value_estimates = [Q1 for action in range(numActions)]
action_counts = [0 for action in range(numActions)]
stp1 = this_observation[
0] # how you convert observation to a number, if state is tabular
atp1 = randInRange(numActions)
action_counts[atp1] += 1
local_action[0] = atp1
last_observation = this_observation # save observation, might be useful on the next step
this_action = local_action
time_step = 1
return this_action
def env_step(
this_action
): # returns (floating point, NumPy array, Boolean), this_action: NumPy array
global local_observation, this_reward_observation #, nStatesSimpleEnv
episode_over = False
#Get a reward from the current action reward distribution
atp1 = int(this_action[0]) # how to extact action
the_reward = randn(bandit_action_values[atp1],
1.0) # rewards drawn from (q*, 1) Gaussian
stp1 = randInRange(
nStatesSimpleEnv) # state transitions are uniform random
#########
local_observation[0] = stp1
this_reward_observation = (the_reward, this_reward_observation[1],
episode_over)
return this_reward_observation