def __init__(self,
environment,
Nsteps=3,
epsilon=0.1,
init_q_val=0.0,
terminal_set=None,
max_steps=sys.maxsize):
self.environment = environment
self.av_coll = ActionValueColl(environment, init_val=init_q_val)
# assume a constant epsilon for now.
self.epsgreedy_obj = EpsilonGreedy(epsilon=epsilon,
const_epsilon=True,
half_life=200,
N_episodes_wo_decay=0)
if terminal_set is None:
self.terminal_set = environment.terminal_set
else:
self.terminal_set = terminal_set
self.Nsteps = Nsteps
self.max_steps = max_steps
self.clear() # initialize the (s,a,r) data structures, t=0, T=inf.
# remember max amount of change due to [s_vector][a_desc]
#delta = np.max( np.absolute( delta_vector ) )
#self.record_changes( s_vector, a_desc, delta )
#return abs(delta) # return the absolute value of change
if __name__ == "__main__":
from introrl.continuous_sims.sim_continuous import ContinuousSimulation
from introrl.continuous_sims.feature_func import FeatureFunction
from introrl.agent_supt.epsilon_calc import EpsilonGreedy
sim = ContinuousSimulation(name='Mountain Car', step_reward=-1.0)
ff = FeatureFunction(sim, name='Proportional', init_w_val=None)
uv = UpdateWVector(ff)
s_vector = sim.get_s_vector()
epsgreedy_obj = EpsilonGreedy(epsilon=0.5,
const_epsilon=True,
half_life=200,
N_episodes_wo_decay=0)
a_best = uv.get_best_eps_greedy_action(s_vector,
epsgreedy_obj=epsgreedy_obj)
print('at', s_vector, ' a_best=', a_best)
print('Best Qsa =', uv.get_max_Qsa(s_vector))
1 if a_desc == 'R' else 0, 1
])
return x_vector
learn_tracker = LearnTracker()
gridworld = get_gridworld(step_reward=-0.1)
NUM_EPISODES = 20000
alpha_obj = Alpha(alpha=0.1)
alpha_obj.set_half_life_for_N_episodes(Nepisodes=NUM_EPISODES,
alpha_final=0.03333333333333)
eps_obj = EpsilonGreedy(epsilon=0.5)
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()
def sarsa_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=1000, 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 SARSA 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 that V(s), action_value_coll, 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_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 SARSA 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 = [ random.choice( environment.limited_start_state_list() ) ]
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 start_hash got to 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
a_desc = action_value_coll.get_best_eps_greedy_action( s_hash, epsgreedy_obj=eg )
for n_episode_steps in range( max_episode_steps ):
# 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')
break
else:
an_desc = action_value_coll.get_best_eps_greedy_action( sn_hash,
epsgreedy_obj=eg )
action_value_coll.sarsa_update( s_hash=s_hash, a_desc=a_desc,
alpha=alpha_obj(), gamma=gamma,
sn_hash=sn_hash, an_desc=an_desc,
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
a_desc = an_desc
# 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:
#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)', end=' ')
print(out_str, end=' ')
print( 'Nterminal episodes =', len(Nterminal_episodes),' of ', len(loop_stateL))
progress_str = out_str
#print()
policy = action_value_coll.get_policy()
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
def __init__(
self,
environment,
learn_tracker=None, # track progress of learning
sa_linear_function=None, # if input, use it.
update_type='sarsa', # can be 'sarsa', 'qlearn'
read_pickle_file='',
save_pickle_file='',
do_summ_print=True,
show_last_change=True,
pcent_progress_print=10,
show_banner=True,
gamma=0.9,
iteration_prints=0,
max_episode_steps=sys.maxsize,
epsilon=0.1, # can be constant or EpsilonGreedy object
alpha=0.1): # can be constant or Alpha object
"""
... GIVEN AN ENVIRONMENT ...
Use basic SARSA or Qlearning algorithm to solve for linear approximation of
STATE-ACTION VALUES, Q(s,a)
Each action is forced to be a DETERMINISTIC action leading to one state and reward.
(If the next state or reward changes, only the new values will be considered)
attribute: self.action_value_linfunc is the linear approximation, Q(s,a) object
A DETERMINISTIC policy can be created externally from the self.action_value_linfunc attribute.
"""
self.environment = environment
self.learn_tracker = learn_tracker
self.save_pickle_file = save_pickle_file
self.do_summ_print = do_summ_print
self.show_last_change = show_last_change
self.pcent_progress_print = pcent_progress_print
self.gamma = gamma
self.iteration_prints = iteration_prints
self.max_episode_steps = max_episode_steps
self.num_episodes = 0
self.num_updates = 0
# if input epsilon is a float, use it to create an EpsilonGreedy object
if type(epsilon) == type(0.1):
self.epsilon_obj = EpsilonGreedy(epsilon=epsilon,
const_epsilon=True)
else:
self.epsilon_obj = epsilon
# if input alpha is a float, use it to create an Alpha object
if type(alpha) == type(0.1):
self.alpha_obj = Alpha(alpha=alpha, const_alpha=True)
else:
self.alpha_obj = alpha
# create the action_value_linfunc for the environment.
self.action_value_linfunc = sa_linear_function
self.update_type = update_type
if read_pickle_file:
self.action_value_linfunc.init_from_pickle_file(read_pickle_file)
if do_summ_print:
print(
'================== EPSILON GREEDY DEFINED AS ========================'
)
self.epsilon_obj.summ_print()
print(
'================== LEARNING RATE DEFINED AS ========================'
)
self.alpha_obj.summ_print()
if show_banner:
s = 'Starting a Maximum of %i %s Semi-Gradient Epsilon Greedy Steps/Episode'%(self.max_episode_steps, update_type.upper()) +\
'\nfor "%s" with Gamma = %g, Alpha = %g'%( environment.name, self.gamma, self.alpha_obj() )
banner(s, banner_char='', leftMargin=0, just='center')
class SA_SemiGradAgent(object):
"""
SARSA or Qlearning semi-gradient agent for linear function approximator.
"""
def __init__(
self,
environment,
learn_tracker=None, # track progress of learning
sa_linear_function=None, # if input, use it.
update_type='sarsa', # can be 'sarsa', 'qlearn'
read_pickle_file='',
save_pickle_file='',
do_summ_print=True,
show_last_change=True,
pcent_progress_print=10,
show_banner=True,
gamma=0.9,
iteration_prints=0,
max_episode_steps=sys.maxsize,
epsilon=0.1, # can be constant or EpsilonGreedy object
alpha=0.1): # can be constant or Alpha object
"""
... GIVEN AN ENVIRONMENT ...
Use basic SARSA or Qlearning algorithm to solve for linear approximation of
STATE-ACTION VALUES, Q(s,a)
Each action is forced to be a DETERMINISTIC action leading to one state and reward.
(If the next state or reward changes, only the new values will be considered)
attribute: self.action_value_linfunc is the linear approximation, Q(s,a) object
A DETERMINISTIC policy can be created externally from the self.action_value_linfunc attribute.
"""
self.environment = environment
self.learn_tracker = learn_tracker
self.save_pickle_file = save_pickle_file
self.do_summ_print = do_summ_print
self.show_last_change = show_last_change
self.pcent_progress_print = pcent_progress_print
self.gamma = gamma
self.iteration_prints = iteration_prints
self.max_episode_steps = max_episode_steps
self.num_episodes = 0
self.num_updates = 0
# if input epsilon is a float, use it to create an EpsilonGreedy object
if type(epsilon) == type(0.1):
self.epsilon_obj = EpsilonGreedy(epsilon=epsilon,
const_epsilon=True)
else:
self.epsilon_obj = epsilon
# if input alpha is a float, use it to create an Alpha object
if type(alpha) == type(0.1):
self.alpha_obj = Alpha(alpha=alpha, const_alpha=True)
else:
self.alpha_obj = alpha
# create the action_value_linfunc for the environment.
self.action_value_linfunc = sa_linear_function
self.update_type = update_type
if read_pickle_file:
self.action_value_linfunc.init_from_pickle_file(read_pickle_file)
if do_summ_print:
print(
'================== EPSILON GREEDY DEFINED AS ========================'
)
self.epsilon_obj.summ_print()
print(
'================== LEARNING RATE DEFINED AS ========================'
)
self.alpha_obj.summ_print()
if show_banner:
s = 'Starting a Maximum of %i %s Semi-Gradient Epsilon Greedy Steps/Episode'%(self.max_episode_steps, update_type.upper()) +\
'\nfor "%s" with Gamma = %g, Alpha = %g'%( environment.name, self.gamma, self.alpha_obj() )
banner(s, banner_char='', leftMargin=0, just='center')
def run_episode(self, start_state, iter_sarsn=None):
"""
Run a single episode of SARSA or Qlearning Semi-Gradient algorithm
If iter_sarsn is input, use it instead of action_value_linfunc calculations.
(Note: the start_state should NOT be in terminal_set if iter_sarsn is input.)
"""
# increment episode counters
self.num_episodes += 1
self.epsilon_obj.inc_N_episodes()
self.alpha_obj.inc_N_episodes()
if self.learn_tracker is not None:
self.learn_tracker.add_new_episode()
# do SARSA or Qlearning Semi-Gradient loops until sn_hash in terminal_set
s_hash = start_state
n_steps_in_episode = 1
while s_hash not in self.environment.terminal_set:
if iter_sarsn is None:
# get best epsilon-greedy action
a_desc = self.action_value_linfunc.get_best_eps_greedy_action( \
s_hash, epsgreedy_obj=self.epsilon_obj )
# check for bad action value
if a_desc is None:
print('break for a_desc==None at s_hash=%s' % str(s_hash))
break
# get next state and reward
sn_hash, reward = self.environment.get_action_snext_reward(
s_hash, a_desc)
else:
# retracing an existing episode
s_hash, a_desc, reward, sn_hash = next(iter_sarsn)
if self.learn_tracker is not None:
self.learn_tracker.add_sarsn_to_current_episode(
s_hash, a_desc, reward, sn_hash)
if sn_hash is None:
print('break for sn_hash==None, #steps=', n_steps_in_episode,
' s_hash=%s' % str(s_hash), ' a_desc=%s' % str(a_desc))
break
# do RL update of Q(s,a) value
if self.update_type == 'sarsa':
an_desc = self.action_value_linfunc.get_best_eps_greedy_action( \
sn_hash, epsgreedy_obj=self.epsilon_obj )
self.action_value_linfunc.sarsa_update(s_hash=s_hash,
a_desc=a_desc,
alpha=self.alpha_obj(),
gamma=self.gamma,
sn_hash=sn_hash,
an_desc=an_desc,
reward=reward)
elif self.update_type == 'qlearn':
self.action_value_linfunc.qlearning_update(
s_hash=s_hash,
a_desc=a_desc,
alpha=self.alpha_obj(),
gamma=self.gamma,
sn_hash=sn_hash,
reward=reward)
self.num_updates += 1
# keep a lid on the max number of episode steps.
if n_steps_in_episode >= self.max_episode_steps:
break
# get ready for next loop
n_steps_in_episode += 1
s_hash = sn_hash
#print(n_steps_in_episode, end=' ')
def summ_print(self, long=True): # pragma: no cover
"""Show State objects in sorted state_hash order."""
print('___ Policy Evaluation Agent Summary ___')
print(' Environment = %s' % self.environment.name)
print(' Update Type = %s' % self.update_type)
print(' Number of Episodes = %g' % self.num_episodes)
print(
'================== EPSILON GREEDY FINAL ========================')
self.epsilon_obj.summ_print()
print(
'================== LEARNING RATE FINAL ========================')
self.alpha_obj.summ_print()
name='TilingsInf',
init_w_val=None,
num_tiles=8,
recenter=True,
num_regionsL=[8, 8])
ff.init_from_pickle_file('mcar_' + ff.desc())
updater = UpdateWVector(ff)
NUM_EPISODES = 5000
UPDATE_TYPE = 'sarsa'
GAMMA = 1.0 # 0.99
eg = EpsilonGreedy(epsilon=0.01,
const_epsilon=True,
half_life=200,
N_episodes_wo_decay=0)
eg.set_half_life_for_N_episodes(Nepisodes=NUM_EPISODES,
epsilon_final=0.00001)
#eg = None
ALPHA = 0.05
if hasattr(ff, 'num_tiles'):
alpha_obj = Alpha(alpha=ALPHA / ff.num_tiles)
else:
alpha_obj = Alpha(alpha=ALPHA)
#alpha_obj.set_half_life_for_N_episodes( Nepisodes=NUM_EPISODES, alpha_final=0.001)
ep_lenL = []
for loop_counter in range(NUM_EPISODES):
if __name__ == "__main__": # pragma: no cover
from introrl.mdp_data.simple_grid_world import get_gridworld
from introrl.policy import Policy
from introrl.agent_supt.epsilon_calc import EpsilonGreedy
from introrl.agent_supt.episode_stats import EpisodeStats
gridworld = get_gridworld()
pi = Policy( environment=gridworld )
pi.set_policy_from_piD( gridworld.get_default_policy_desc_dict() )
#pi.summ_print()
eg = EpsilonGreedy(epsilon=0.2, const_epsilon=True, half_life=200,
N_episodes_wo_decay=0)
episode = make_episode( (2,0), pi, gridworld, gridworld.terminal_set, eps_greedy=eg )
episode.summ_print()
print('-'*55)
print(' First "S" Visit Reversed Return Values.')
rev_rtnL = episode.get_rev_discounted_returns( gamma=0.9, first_visit=True, visit_type='S')
for _ in rev_rtnL:
print( _ )
# --------------------------------------------
print('#'*55)
episode = make_episode( (2,0), pi, gridworld, gridworld.terminal_set, eps_greedy=eg, episode=episode )
episode.summ_print()
def mc_epsilon_greedy(
environment,
initial_policy='default', # can be 'default', 'random', policy_dictionary
read_pickle_file='',
save_pickle_file='',
use_list_of_start_states=True, # use list OR single start state of environment.
iter_all_start_actions=False, # pick random or iterate all starting actions
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,
iteration_prints=0,
max_episode_steps=10000,
epsilon=0.1,
const_epsilon=True,
half_life=200,
N_episodes_wo_decay=0):
"""
... GIVEN AN ENVIRONMENT ...
apply Monte Carlo Exploring Starts to find the OPTIMAL POLICY
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.
"""
eps_greedy = EpsilonGreedy(epsilon=epsilon,
const_epsilon=const_epsilon,
half_life=half_life,
N_episodes_wo_decay=N_episodes_wo_decay)
# create Policy and ActionValueRunAveColl objects
policy = Policy(environment=environment)
if initial_policy == 'default':
print('Initializing Policy to "default" in mc_epsilon_greedy')
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_epsilon_greedy')
policy.intialize_policy_to_random(env=environment)
else:
print('Initializing Policy to "custom policy" in mc_epsilon_greedy')
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='*')
print(
'================== EPSILON GREEDY DEFINED AS ========================'
)
eps_greedy.summ_print()
s = 'Starting a Maximum of %i Monte Carlo Epsilon Greedy\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
limited_start_stateL = environment.limited_start_state_list()
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
Nterminal_episodes = set()
# Iterate over a list of known possible start states
if use_list_of_start_states:
loop_stateL = limited_start_stateL
random.shuffle(loop_stateL)
else:
#loop_stateL = [ random.choice( limited_start_stateL ) ]
loop_stateL = [environment.start_state_hash]
for start_hash in loop_stateL:
if iter_all_start_actions: # Iterate over ALL ACTIONS of start_hash
a_descL = environment.get_state_legal_action_list(start_hash)
else:
a_desc = policy.get_single_action(start_hash)
# if not iterating all actions, make sure first action has eps-greedy applied
a_desc = eps_greedy(
a_desc,
environment.get_state_legal_action_list(start_hash))
a_descL = [a_desc]
# randomize action order
random.shuffle(a_descL)
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=eps_greedy)
eps_greedy.inc_N_episodes()
num_episodes += 1
if episode.is_done():
Nterminal_episodes.add(start_hash)
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)',
end=' ')
print('Nterminal episodes =', len(Nterminal_episodes), ' of ',
len(loop_stateL))
progress_str = out_str
print()
if do_summ_print:
s = ''
if num_episodes >= max_num_episodes:
s = ' (NOTE: STOPPED ON MAX-ITERATIONS)'
print('Exited Epsilon Greedy, MC First-Visit Value Iterations', 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)
print('Nterminal episodes =', len(Nterminal_episodes), ' of ',
len(loop_stateL))
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 __init__(
self,
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='',
do_summ_print=True,
show_last_change=True,
pcent_progress_print=10,
show_banner=True,
gamma=0.9,
iteration_prints=0,
max_episode_steps=sys.maxsize,
epsilon=0.1, # can be constant or EpsilonGreedy object
alpha=0.1): # can be constant or Alpha object
"""
... GIVEN AN ENVIRONMENT ...
Use basic Dyna-Q algorithm to solve for STATE-ACTION VALUES, Q(s,a)
Each action is forced to be a DETERMINISTIC action leading to one state and reward.
(If the next state or reward changes, only the new values will be considered)
attribute: self.action_value_coll is the ActionValueColl, Q(s,a) object
A DETERMINISTIC policy can be created externally from the self.action_value_coll attribute.
"""
self.environment = environment
self.learn_tracker = learn_tracker
self.save_pickle_file = save_pickle_file
self.do_summ_print = do_summ_print
self.show_last_change = show_last_change
self.pcent_progress_print = pcent_progress_print
self.gamma = gamma
self.iteration_prints = iteration_prints
self.max_episode_steps = max_episode_steps
self.num_episodes = 0
self.num_updates = 0
# if input epsilon is a float, use it to create an EpsilonGreedy object
if type(epsilon) == type(0.1):
self.epsilon_obj = EpsilonGreedy(epsilon=epsilon,
const_epsilon=True)
else:
self.epsilon_obj = epsilon
# if input alpha is a float, use it to create an Alpha object
if type(alpha) == type(0.1):
self.alpha_obj = Alpha(alpha=alpha, const_alpha=True)
else:
self.alpha_obj = alpha
# create the action_value_coll for the environment.
if initial_action_value_coll is None:
self.action_value_coll = ActionValueColl(environment,
init_val=initial_Qsa)
else:
self.action_value_coll = initial_action_value_coll
if read_pickle_file:
self.action_value_coll.init_from_pickle_file(read_pickle_file)
# initialize the model that will build from experience
# do not build full model description on Model init, states not visited
# by the RL portion will have no returns values.
self.model = Model(environment, build_initial_model=False)
#for s_hash, aD in self.action_value_coll.QsaD.items():
# for a_desc, Q in aD.items():
# self.model.add_action( s_hash, a_desc )
if do_summ_print:
print(
'================== EPSILON GREEDY DEFINED AS ========================'
)
self.epsilon_obj.summ_print()
print(
'================== LEARNING RATE DEFINED AS ========================'
)
self.alpha_obj.summ_print()
if show_banner:
s = 'Starting a Maximum of %i Dyna-Q Epsilon Greedy Steps/Episode'%self.max_episode_steps +\
'\nfor "%s" with Gamma = %g, Alpha = %g'%( environment.name, self.gamma, self.alpha_obj() )
banner(s, banner_char='', leftMargin=0, just='center')
class DynaQAgent(object):
"""
DynaQ Agent.
"""
def __init__(
self,
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='',
do_summ_print=True,
show_last_change=True,
pcent_progress_print=10,
show_banner=True,
gamma=0.9,
iteration_prints=0,
max_episode_steps=sys.maxsize,
epsilon=0.1, # can be constant or EpsilonGreedy object
alpha=0.1): # can be constant or Alpha object
"""
... GIVEN AN ENVIRONMENT ...
Use basic Dyna-Q algorithm to solve for STATE-ACTION VALUES, Q(s,a)
Each action is forced to be a DETERMINISTIC action leading to one state and reward.
(If the next state or reward changes, only the new values will be considered)
attribute: self.action_value_coll is the ActionValueColl, Q(s,a) object
A DETERMINISTIC policy can be created externally from the self.action_value_coll attribute.
"""
self.environment = environment
self.learn_tracker = learn_tracker
self.save_pickle_file = save_pickle_file
self.do_summ_print = do_summ_print
self.show_last_change = show_last_change
self.pcent_progress_print = pcent_progress_print
self.gamma = gamma
self.iteration_prints = iteration_prints
self.max_episode_steps = max_episode_steps
self.num_episodes = 0
self.num_updates = 0
# if input epsilon is a float, use it to create an EpsilonGreedy object
if type(epsilon) == type(0.1):
self.epsilon_obj = EpsilonGreedy(epsilon=epsilon,
const_epsilon=True)
else:
self.epsilon_obj = epsilon
# if input alpha is a float, use it to create an Alpha object
if type(alpha) == type(0.1):
self.alpha_obj = Alpha(alpha=alpha, const_alpha=True)
else:
self.alpha_obj = alpha
# create the action_value_coll for the environment.
if initial_action_value_coll is None:
self.action_value_coll = ActionValueColl(environment,
init_val=initial_Qsa)
else:
self.action_value_coll = initial_action_value_coll
if read_pickle_file:
self.action_value_coll.init_from_pickle_file(read_pickle_file)
# initialize the model that will build from experience
# do not build full model description on Model init, states not visited
# by the RL portion will have no returns values.
self.model = Model(environment, build_initial_model=False)
#for s_hash, aD in self.action_value_coll.QsaD.items():
# for a_desc, Q in aD.items():
# self.model.add_action( s_hash, a_desc )
if do_summ_print:
print(
'================== EPSILON GREEDY DEFINED AS ========================'
)
self.epsilon_obj.summ_print()
print(
'================== LEARNING RATE DEFINED AS ========================'
)
self.alpha_obj.summ_print()
if show_banner:
s = 'Starting a Maximum of %i Dyna-Q Epsilon Greedy Steps/Episode'%self.max_episode_steps +\
'\nfor "%s" with Gamma = %g, Alpha = %g'%( environment.name, self.gamma, self.alpha_obj() )
banner(s, banner_char='', leftMargin=0, just='center')
def run_episode(self, start_state, Nplanning_loops=5, iter_sarsn=None):
"""
Run a single episode of Dyna-Q algorithm
If iter_sarsn is input, use it instead of action_value_coll calculations.
(Note: the start_state should NOT be in terminal_set if iter_sarsn is input.)
"""
# increment episode counters
self.num_episodes += 1
self.epsilon_obj.inc_N_episodes()
self.alpha_obj.inc_N_episodes()
if self.learn_tracker is not None:
self.learn_tracker.add_new_episode()
# do dyna_q loops until sn_hash in terminal_set
s_hash = start_state
n_steps_in_episode = 1
while s_hash not in self.environment.terminal_set:
if iter_sarsn is None:
# get best epsilon-greedy action
a_desc = self.action_value_coll.get_best_eps_greedy_action( \
s_hash, epsgreedy_obj=self.epsilon_obj )
# check for bad action value
if a_desc is None:
print('break for a_desc==None at s_hash=%s' % str(s_hash))
break
# get next state and reward
sn_hash, reward = self.environment.get_action_snext_reward(
s_hash, a_desc)
else:
# retracing an existing episode
s_hash, a_desc, reward, sn_hash = next(iter_sarsn)
if self.learn_tracker is not None:
self.learn_tracker.add_sarsn_to_current_episode(
s_hash, a_desc, reward, sn_hash)
if sn_hash is None:
print('break for sn_hash==None, #steps=', n_steps_in_episode,
' s_hash=%s' % str(s_hash), ' a_desc=%s' % str(a_desc))
break
# do RL update of Q(s,a) value
self.action_value_coll.qlearning_update(s_hash=s_hash,
a_desc=a_desc,
sn_hash=sn_hash,
alpha=self.alpha_obj(),
gamma=self.gamma,
reward=reward)
self.num_updates += 1
# give the above experience to the model
self.model.add_action(s_hash, a_desc)
# force DETERMINISTIC next state and reward.
self.model.save_deterministic_action_results(s_hash,
a_desc,
sn_hash,
reward_val=reward)
# do NOT use simple save_action_results... it allows NON-DETERMINISTIC next state.
#self.model.save_action_results( s_hash, a_desc, sn_hash, reward_val=reward)
# --------------------------------- Planning Loop ------------------------
# make Nplanning_loops calls to model
for n_plan in range(Nplanning_loops):
s_model = self.model.get_random_state()
#print(s_model, end=' ')
# vanilla DynaQ
a_model = self.model.get_random_action(s_model)
#sn_model, r_model = self.environment.get_action_snext_reward( s_model, a_model )
sn_model, r_model = self.model.get_sample_sn_r(
s_model, a_model)
# update for the DynaQ results.
self.action_value_coll.qlearning_update(s_hash=s_model,
a_desc=a_model,
sn_hash=sn_model,
alpha=self.alpha_obj(),
gamma=self.gamma,
reward=r_model)
self.num_updates += 1
# keep a lid on the max number of episode steps.
if n_steps_in_episode >= self.max_episode_steps:
break
# get ready for next loop
n_steps_in_episode += 1
s_hash = sn_hash
#print(n_steps_in_episode, end=' ')
def summ_print(self, long=True): # pragma: no cover
"""Show State objects in sorted state_hash order."""
print('___ Policy Evaluation Agent Summary ___')
print(' Environment = %s' % self.environment.name)
print(' Number of Episodes = %g' % self.num_episodes)
