def get_policy(self):
policy = Policy(environment=self.environment)
for s_hash in self.environment.iter_all_action_states():
a_desc = self.get_best_greedy_action(s_hash)
policy.set_sole_action(s_hash, a_desc)
return policy
import matplotlib.pyplot as plt
from introrl.td_funcs.td0_prediction import td0_prediction
from introrl.utils.running_ave import RunningAve
from introrl.mc_funcs.mc_ev_prediction import mc_every_visit_prediction
from introrl.policy import Policy
from introrl.agent_supt.state_value_coll import StateValueColl
from introrl.agent_supt.nstep_td_eval_walker import NStepTDWalker
from introrl.mdp_data.random_walk_generic_mrp import get_random_walk
from introrl.agent_supt.episode_maker import make_episode
GAMMA=1.0
AVE_OVER = 100
rw_mrp = get_random_walk(Nside_states=9, win_reward=1.0, lose_reward=-1.0, step_reward=0.0)
policy = Policy( environment=rw_mrp )
policy.intialize_policy_to_equiprobable() # should be equiprobable from above init already
episode_obj = make_episode( 'C', policy, rw_mrp )
fig, ax = plt.subplots()
# ---------------- set up true value data for RMS calc --------------------
true_valueD = {'C':0.0} # { 'Win':0.0, 'Lose':0.0}
#print('rw_mrp.get_num_states() = ',rw_mrp.get_num_states())
delta = 2.0 / (rw_mrp.get_num_states()-1)
Nsides = int( rw_mrp.get_num_states() / 2) - 1
d = 0.0
for i in range(1, Nsides+1 ):
eps_obj.set_half_life_for_N_episodes(Nepisodes=NUM_EPISODES,
epsilon_final=0.16666666666666)
agent = SA_SemiGradAgent(environment=gridworld,
update_type='qlearn',
sa_linear_function=LazyProgrammerMaze(gridworld),
learn_tracker=learn_tracker,
gamma=0.9,
alpha=alpha_obj,
epsilon=eps_obj)
for i in range(NUM_EPISODES):
agent.run_episode((2, 0))
print()
agent.summ_print()
print('-' * 77)
#learn_tracker.summ_print()
#print('-'*77)
agent.action_value_linfunc.summ_print(fmt_Q='%.4f')
print('-' * 77)
policy = Policy(environment=gridworld)
for s_hash in gridworld.iter_all_action_states():
a_desc = agent.action_value_linfunc.get_best_eps_greedy_action(
s_hash, epsgreedy_obj=None)
policy.set_sole_action(s_hash, a_desc)
policy.summ_print(environment=gridworld, verbosity=0)
from introrl.black_box_sims.random_walk_1000 import RandomWalk_1000Simulation
from introrl.agent_supt.episode_maker import make_episode
from introrl.policy import Policy
NUM_EPISODES = 100000
countD = {} # index=state, value=count
RW = RandomWalk_1000Simulation()
policy = Policy(environment=RW)
policy.intialize_policy_to_equiprobable( env=RW )
for Nepi in range(NUM_EPISODES):
episode = make_episode(500, policy, RW, max_steps=10000)
for dr in episode.get_rev_discounted_returns( gamma=1.0 ):
(s_hash, a_desc, reward, sn_hash, G) = dr
countD[ s_hash ] = countD.get( s_hash, 0 ) + 1
SUM_VISITS = sum( list(countD.values()) )
freqL = []
for i in range(1,1001):
freqL.append( countD.get(i,0) / float(SUM_VISITS) )
# copy and paste list into plot script
print('freqL =', repr(freqL))
from introrl.dp_funcs.dp_policy_iter import dp_policy_iteration
from introrl.policy import Policy
from introrl.state_values import StateValues
from introrl.mdp_data.car_rental import get_env
from introrl.utils import pickle_esp
env = get_env()
policy = Policy(environment=env)
policy.intialize_policy_to_random(env=env)
state_value = StateValues(env)
state_value.init_Vs_to_zero()
dp_policy_iteration(policy,
state_value,
do_summ_print=True,
show_start_policy=True,
max_iter=1000,
err_delta=0.0001,
gamma=0.9)
diag_colorD = {
'5': 'r',
'4': 'g',
'3': 'b',
'2': 'c',
'1': 'y',
'0': 'w',
'-5': 'r',
'-4': 'g',
def qlearning_epsilon_greedy(
environment,
learn_tracker=None, # track progress of learning
initial_Qsa=0.0, # init non-terminal_set of V(s) (terminal_set=0.0)
initial_action_value_coll=None, # if input, use it.
read_pickle_file='',
save_pickle_file='',
use_list_of_start_states=False, # use list OR single start state of environment.
do_summ_print=True,
show_last_change=True,
fmt_Q='%g',
fmt_R='%g',
pcent_progress_print=10,
show_banner=True,
max_num_episodes=sys.maxsize,
min_num_episodes=10,
max_abserr=0.001,
gamma=0.9,
iteration_prints=0,
max_episode_steps=sys.maxsize,
epsilon=0.1,
const_epsilon=True,
epsilon_half_life=200,
alpha=0.1,
const_alpha=True,
alpha_half_life=200,
N_episodes_wo_decay=0):
"""
... GIVEN AN ENVIRONMENT ...
apply Q-Learning Temporal Difference to find the OPTIMAL POLICY and STATE VALUES
Returns: Policy and ActionValueColl objects
Use Episode Discounted Returns to find V(s), State-Value Function
Terminates when abserr < max_abserr
Assume environment attached to policy will have method "get_any_action_state_hash"
in order to begin at any action state.
CREATES BOTH policy AND action_value_coll OBJECTS.
"""
# create EpsilonGreedy, Alpha and ActionValueColl objects
eg = EpsilonGreedy(epsilon=epsilon,
const_epsilon=const_epsilon,
half_life=epsilon_half_life,
N_episodes_wo_decay=N_episodes_wo_decay)
alpha_obj = Alpha(alpha=alpha,
const_alpha=const_alpha,
half_life=alpha_half_life)
if initial_action_value_coll is None:
action_value_coll = ActionValueColl(environment, init_val=initial_Qsa)
else:
action_value_coll = initial_action_value_coll
#action_value_coll.summ_print()
num_s_hash = len(environment.get_all_action_state_hashes())
if read_pickle_file:
action_value_coll.init_from_pickle_file(read_pickle_file)
if do_summ_print:
print(
'================== EPSILON GREEDY DEFINED AS ========================'
)
eg.summ_print()
print(
'================== LEARNING RATE DEFINED AS ========================'
)
alpha_obj.summ_print()
if show_banner:
s = 'Starting a Maximum of %i Q-Learning Epsilon Greedy Episodes'%max_num_episodes +\
'\nfor "%s" with Gamma = %g, Alpha = %g'%( environment.name, gamma, alpha_obj() )
banner(s, banner_char='', leftMargin=0, just='center')
# Iterate over a list of known possible start states
if use_list_of_start_states:
loop_stateL = environment.limited_start_state_list()
else:
loop_stateL = [environment.start_state_hash]
if show_banner:
print(
'======================= Iterating over Start States =================================='
)
print(loop_stateL)
print(
'======================================================================================'
)
# set counter and flag
episode_loop_counter = 0
keep_looping = True
progress_str = ''
while (episode_loop_counter <= max_num_episodes - 1) and keep_looping:
keep_looping = False
abserr = 0.0 # calculated below as part of termination criteria
Nterminal_episodes = set(
) # tracks if ended at terminal_set or max_num_episodes
for start_hash in loop_stateL:
episode_loop_counter += 1
if episode_loop_counter > max_num_episodes:
break
if learn_tracker is not None:
learn_tracker.add_new_episode()
s_hash = start_hash
for n_episode_steps in range(max_episode_steps):
a_desc = action_value_coll.get_best_eps_greedy_action(
s_hash, epsgreedy_obj=eg)
# Begin an episode
if a_desc is None:
Nterminal_episodes.add(start_hash)
print('break for a_desc==None')
break
else:
sn_hash, reward = environment.get_action_snext_reward(
s_hash, a_desc)
if learn_tracker is not None:
learn_tracker.add_sarsn_to_current_episode(
s_hash, a_desc, reward, sn_hash)
if sn_hash is None:
Nterminal_episodes.add(start_hash)
print('break for sn_hash==None, #steps=',
n_episode_steps, ' s_hash=%s' % str(s_hash),
' a_desc=%s' % str(a_desc))
break
else:
action_value_coll.qlearning_update(s_hash=s_hash,
a_desc=a_desc,
sn_hash=sn_hash,
alpha=alpha_obj(),
gamma=gamma,
reward=reward)
if sn_hash in environment.terminal_set:
Nterminal_episodes.add(start_hash)
if (n_episode_steps == 0) and (num_s_hash > 2):
print(
'1st step break for sn_hash in terminal_set',
sn_hash, ' s_hash=%s' % str(s_hash),
' a_desc=%s' % str(a_desc))
break
s_hash = sn_hash
# increment episode counter on EpsilonGreedy and Alpha objects
eg.inc_N_episodes()
alpha_obj.inc_N_episodes()
abserr = action_value_coll.get_biggest_action_state_err()
if abserr > max_abserr:
keep_looping = True
if episode_loop_counter < min_num_episodes:
keep_looping = True # must loop for min_num_episodes at least
pc_done = 100.0 * float(episode_loop_counter) / float(max_num_episodes)
if pcent_progress_print > 0:
out_str = '%3i%%' % (pcent_progress_print *
(int(pc_done / float(pcent_progress_print))))
else:
out_str = progress_str
if out_str != progress_str:
print(out_str, end=' ')
print('Nterminal episodes =', len(Nterminal_episodes), ' of ',
len(loop_stateL))
progress_str = out_str
policy = Policy(environment=environment)
for s_hash in environment.iter_all_action_states():
a_desc = action_value_coll.get_best_eps_greedy_action(
s_hash, epsgreedy_obj=None)
policy.set_sole_action(s_hash, a_desc)
if do_summ_print:
s = ''
if episode_loop_counter >= max_num_episodes:
s = ' (NOTE: STOPPED ON MAX-ITERATIONS)'
print('Exited Epsilon Greedy, TD(0) Value Iteration', s)
print(' # episodes =', episode_loop_counter,
' (min limit=%i)' % min_num_episodes,
' (max limit=%i)' % max_num_episodes)
print(' gamma =', gamma)
print(' estimated err =', abserr)
print(' Error limit =', max_abserr)
print('Nterminal episodes =', len(Nterminal_episodes), ' of ',
len(loop_stateL))
action_value_coll.summ_print(show_last_change=show_last_change,
fmt_Q=fmt_Q)
policy.summ_print(environment=environment,
verbosity=0,
show_env_states=False)
try: # sims may not have a layout_print
environment.layout_print(vname='reward',
fmt=fmt_R,
show_env_states=False,
none_str='*')
except:
pass
print(
'================== EPSILON GREEDY DEFINED AS ========================'
)
eg.summ_print()
print(
'================== LEARNING RATE DEFINED AS ========================'
)
alpha_obj.summ_print()
if save_pickle_file:
policy.save_to_pickle_file(save_pickle_file)
action_value_coll.save_to_pickle_file(save_pickle_file)
return policy, action_value_coll #, steps_per_episodeL, reward_sum_per_episodeL
print('_____________ Value Iteration ________________')
else:
print('_____________ Policy Iteration ________________')
for gamma in (0.001, 0.01, 0.1, 0.25, 0.5, 0.75, 0.9, 0.95, 0.99, 0.999):
if do_VI:
policy, sv = dp_value_iteration(robot,
do_summ_print=False,
fmt_V='%.1f',
max_iter=1000,
err_delta=0.001,
gamma=gamma)
else:
policy = Policy(environment=robot)
policy.set_policy_from_piD(robot.get_default_policy_desc_dict())
sv = StateValues(robot)
sv.init_Vs_to_zero()
dp_policy_iteration(policy,
sv,
do_summ_print=False,
max_iter=1000,
err_delta=0.001,
gamma=gamma)
print('gamma=%5g' % gamma, ' Fallen=', policy.get_single_action('Fallen'),
' Moving=', policy.get_single_action('Moving'), ' Standing=',
policy.get_single_action('Standing'), ' Fallen=',
from introrl.mc_funcs.mc_fv_prediction import mc_first_visit_prediction
from introrl.black_box_sims.blackjack_sim import BlackJackSimulation
from introrl.policy import Policy
from introrl.agent_supt.state_value_run_ave_coll import StateValueRunAveColl
BJ = BlackJackSimulation()
pi = Policy(environment=BJ)
# default policy is hit on everything except 20 & 21.
pi.set_policy_from_piD(BJ.get_default_policy_desc_dict())
sv = StateValueRunAveColl(BJ)
if 1:
mc_first_visit_prediction(pi,
sv,
max_num_episodes=10000,
max_abserr=0.001,
gamma=1.0)
sv.save_to_pickle_file(fname='mc_blackjack_10000_eval')
else:
mc_first_visit_prediction(pi,
sv,
max_num_episodes=500000,
max_abserr=0.001,
gamma=1.0)
sv.save_to_pickle_file(fname='mc_blackjack_500000_eval')
from introrl.policy import Policy
from introrl.black_box_sims.racetrack_1_sim import RaceTrack_1
import matplotlib
import matplotlib.pyplot as plt
import matplotlib.patches as mpatches
RT = RaceTrack_1()
#pi = Policy( environment=RT )
policyD = Policy().read_pickle_file('racetrack_1_sim')
#pi.init_from_pickle_file( 'racetrack_1_sim' )
fig, ax = plt.subplots()
for (j, i) in RT.racetrack_area:
rect = mpatches.Rectangle((i - .5, j - .5),
1.0,
1.0,
ec="none",
color='blue',
alpha=0.3)
ax.add_patch(rect)
for (j, i, _, _) in RT.starting_lineL:
rect = mpatches.Rectangle((i - .5, j - .5),
1.0,
1.0,
ec="none",
color='yellow',
alpha=1.)
def setUp(self):
unittest.TestCase.setUp(self)
self.gridworld = get_gridworld()
self.P = Policy(environment=self.gridworld)
self.P.intialize_policy_to_equiprobable(env=self.gridworld)
if 1:
pi, av = mc_epsilon_greedy(RT,
initial_policy='default',
first_visit=True,
do_summ_print=False,
showRunningAve=False,
fmt_Q='%g',
fmt_R='%g',
show_initial_policy=False,
max_num_episodes=1000,
min_num_episodes=10,
max_abserr=0.001,
gamma=0.9,
iteration_prints=0,
max_episode_steps=10000,
epsilon=0.1,
const_epsilon=True,
half_life=500,
N_episodes_wo_decay=0)
pi.save_to_pickle_file('racetrack_2_sim')
else:
pi = Policy(environment=RT)
pi.init_from_pickle_file('racetrack_2_sim')
fig, ax = plt.subplots()
RT.plot_policy(ax, pi)
plt.show()
fig.savefig("racetrack_2_sim.png")
def mc_exploring_starts(environment,
initial_policy='default',
read_pickle_file='',
save_pickle_file='',
first_visit=True,
do_summ_print=True,
showRunningAve=False,
fmt_Q='%g',
fmt_R='%g',
show_initial_policy=True,
max_num_episodes=1000,
min_num_episodes=10,
max_abserr=0.001,
gamma=0.9,
max_episode_steps=10000,
iteration_prints=0):
"""
... GIVEN AN ENVIRONMENT ...
apply Monte Carlo Exploring Starts to find the OPTIMAL POLICY
initial_policy can be 'default', 'random', policy_dictionary, Policy object
Returns: Policy and ActionValueRunAveColl objects
Use Episode Discounted Returns to find Q(s,a), Action-Value Function
Terminates when abserr < max_abserr
Assume that Q(s,a), action_value_ave, has been initialized prior to call.
Assume environment attached to policy will have method "get_any_action_state_hash"
in order to begin at any action state.
CREATES BOTH policy AND action_value OBJECTS.
"""
# create Policy and ActionValueRunAveColl objects
policy = Policy(environment=environment)
if initial_policy == 'default':
print('Initializing Policy to "default" in mc_exploring_starts')
policy.learn_a_legal_action_from_env(env=environment)
policy.set_policy_from_piD(environment.get_default_policy_desc_dict())
elif initial_policy == 'random':
print('Initializing Policy to "random" in mc_exploring_starts')
policy.intialize_policy_to_random(env=environment)
elif isinstance(initial_policy, Policy):
policy = initial_policy
else:
print('Initializing Policy to "custom policy" in mc_exploring_starts')
policy.learn_a_legal_action_from_env(env=environment)
policy.set_policy_from_piD(initial_policy)
action_value_ave = ActionValueRunAveColl(environment)
action_value_ave.init_Qsa_to_zero(
) # Terminal states w/o an action are NOT included
#action_value_ave.summ_print()
if read_pickle_file:
policy.init_from_pickle_file(read_pickle_file)
action_value_ave.init_from_pickle_file(read_pickle_file)
if do_summ_print:
if show_initial_policy:
print(
'=============== STARTING WITH THE INITIAL POLICY ===================='
)
policy.summ_print(verbosity=0,
environment=environment,
show_env_states=False,
none_str='*')
s = 'Starting a Maximum of %i Monte Carlo Exploring Start Episodes\nfor "%s" with Gamma = %g'%\
(max_num_episodes, environment.name, gamma)
banner(s, banner_char='', leftMargin=0, just='center')
# create an Episode object for getting returns
episode = Episode(environment.name + ' Episode')
# set counter and flag
num_episodes = 0
keep_looping = True
progress_str = ''
while (num_episodes <= max_num_episodes - 1) and keep_looping:
keep_looping = False
abserr = 0.0 # calculated below as part of termination criteria
for start_hash in environment.iter_all_action_states(randomize=True):
a_descL = environment.get_state_legal_action_list(start_hash)
# randomize action order
random.shuffle(a_descL)
# try every initial action for each start_hash
for a_desc in a_descL:
# break from inner loop if max_num_episodes is hit.
if num_episodes >= max_num_episodes:
break
make_episode(start_hash,
policy,
environment,
environment.terminal_set,
episode=episode,
first_a_desc=a_desc,
max_steps=max_episode_steps,
eps_greedy=None)
num_episodes += 1
for dr in episode.get_rev_discounted_returns(
gamma=gamma, first_visit=first_visit, visit_type='SA'):
# look at each step from episode and calc average Q(s,a)
(s, a, r, sn, G) = dr
action_value_ave.add_val(s, a, G)
aL = environment.get_state_legal_action_list(s)
if aL:
best_a_desc, best_a_val = aL[0], float('-inf')
bestL = [best_a_desc]
for a in aL:
q = action_value_ave.get_ave(s, a)
if q > best_a_val:
best_a_desc, best_a_val = a, q
bestL = [a]
elif q == best_a_val:
bestL.append(a)
best_a_desc = random.choice(bestL)
policy.set_sole_action(s, best_a_desc)
abserr = action_value_ave.get_biggest_action_state_err()
if abserr > max_abserr:
keep_looping = True
if num_episodes < min_num_episodes:
keep_looping = True # must loop for min_num_episodes at least
pc_done = 100.0 * float(num_episodes) / float(max_num_episodes)
out_str = '%3i%%' % (5 * (int(pc_done / 5.0)))
if out_str != progress_str:
score = environment.get_policy_score(policy=policy,
start_state_hash=None,
step_limit=1000)
print(out_str, ' score=%s' % str(score),
' = (r_sum, n_steps, msg)', ' estimated err =', abserr)
progress_str = out_str
if do_summ_print:
s = ''
if num_episodes >= max_num_episodes:
s = ' (NOTE: STOPPED ON MAX-ITERATIONS)'
print('Exited MC First-Visit Value Iteration', s)
print(' num episodes =', num_episodes,
' (min limit=%i)' % min_num_episodes,
' (max limit=%i)' % max_num_episodes)
print(' gamma =', gamma)
print(' estimated err =', abserr)
print(' Error limit =', max_abserr)
action_value_ave.summ_print(showRunningAve=showRunningAve, fmt_Q=fmt_Q)
policy.summ_print(environment=environment,
verbosity=0,
show_env_states=False)
try: # sims may not have a layout_print
environment.layout_print(vname='reward',
fmt=fmt_R,
show_env_states=False,
none_str='*')
except:
pass
if save_pickle_file:
policy.save_to_pickle_file(save_pickle_file)
action_value_ave.save_to_pickle_file(save_pickle_file)
return policy, action_value_ave
def dp_value_iteration(environment,
allow_multi_actions=False,
do_summ_print=True,
fmt_V='%g',
fmt_R='%g',
max_iter=1000,
err_delta=0.001,
gamma=0.9,
iteration_prints=0):
"""
... GIVEN AN ENVIRONMENT ...
apply Value Iteration to find the OPTIMAL POLICY
Returns: policy and state_value objects
Terminates when delta < err_delta * VI_STOP_CRITERIA
CREATES BOTH policy AND state_value OBJECTS.
If allow_multi_actions is True, policy will include all actions
within err_delta of best action.
"""
# create Policy and StateValues objects
policy = Policy(environment=environment)
policy.intialize_policy_to_random(env=environment)
state_value = StateValues(environment)
state_value.init_Vs_to_zero() # Terminal states need to be 0.0
#state_value.summ_print()
# set counter and flag
loop_counter = 0
all_done = False
# value-iteration stopping criteria
# if gamme==1.0 value iteration will never stop SO limit to gamma==0.999 stop criteria
# (VI terminates if delta < err_delta * VI_STOP_CRITERIA)
# (typically err_delta = 0.001)
VI_STOP_CRITERIA = max((1.0 - gamma) / gamma, (1.0 - 0.999) / 0.999)
error_limit = err_delta * VI_STOP_CRITERIA
while (loop_counter < max_iter) and (not all_done):
loop_counter += 1
all_done = True
delta = 0.0 # used to calc largest change in state_value
for s_hash in policy.iter_all_policy_states():
VsD = {
} # will hold: index=a_desc, value=V(s) for all transitions of a_desc from s_hash
# MUST include currently zero prob actions
for a_desc, a_prob in policy.iter_policy_ap_for_state(
s_hash, incl_zero_prob=True):
calcd_v = 0.0
for sn_hash, t_prob, reward in \
environment.iter_next_state_prob_reward(s_hash, a_desc, incl_zero_prob=False):
calcd_v += t_prob * (reward + gamma * state_value(sn_hash))
VsD[a_desc] = calcd_v
best_a_desc, best_a_val = argmax_vmax_dict(VsD)
delta = max(delta, abs(best_a_val - state_value(s_hash)))
state_value[s_hash] = best_a_val
if delta > error_limit:
all_done = False
if iteration_prints and (loop_counter % iteration_prints == 0):
print('Loop:%6i' % loop_counter, ' delta=%g' % delta)
# Now that State-Values have been determined, set policy
for s_hash in policy.iter_all_policy_states():
VsD = {
} # will hold: index=a_desc, value=V(s) for all transitions of a_desc from s_hash
# MUST include zero prob actions
for a_desc, a_prob in policy.iter_policy_ap_for_state(
s_hash, incl_zero_prob=True):
calcd_v = 0.0
for sn_hash, t_prob, reward in \
environment.iter_next_state_prob_reward(s_hash, a_desc, incl_zero_prob=False):
calcd_v += t_prob * (reward + gamma * state_value(sn_hash))
VsD[a_desc] = calcd_v
if allow_multi_actions:
best_a_list, best_a_val = multi_argmax_vmax_dict(
VsD, err_delta=err_delta)
policy.set_sole_action(s_hash,
best_a_list[0]) # zero all other actions
prob = 1.0 / len(best_a_list)
for a_desc in best_a_list:
policy.set_action_prob(s_hash, a_desc, prob=prob)
else:
best_a_desc, best_a_val = argmax_vmax_dict(VsD)
policy.set_sole_action(s_hash, best_a_desc)
if do_summ_print:
s = ''
if loop_counter >= max_iter:
s = ' (NOTE: STOPPED ON MAX-ITERATIONS)'
print('Exited Value Iteration', s)
print(' iterations =', loop_counter, ' (limit=%i)' % max_iter)
print(' measured delta =', delta)
print(' gamma =', gamma)
print(' err_delta =', err_delta)
print(' error limit =', error_limit)
print(' STOP CRITERIA =', VI_STOP_CRITERIA)
state_value.summ_print(fmt_V=fmt_V)
policy.summ_print(environment=environment,
verbosity=0,
show_env_states=False)
environment.layout_print(vname='reward',
fmt=fmt_R,
show_env_states=False,
none_str='*')
return policy, state_value