def init_Qsa_to_zero(self):
# initialize to 0.0 for all states, terminal and non-terminal.
for s_hash in self.environment.iter_all_states():
if s_hash not in self.Qsa_RaveD:
self.Qsa_RaveD[s_hash] = {}
# may not be any actions in terminal state, so set None action.
if s_hash in self.environment.terminal_set:
self.Qsa_RaveD[s_hash][None] = RunningAve(name=str(s_hash) +
' None')
aL = self.environment.get_state_legal_action_list(s_hash)
for a_desc in aL:
self.Qsa_RaveD[s_hash][a_desc] = RunningAve(name=str(s_hash) +
' ' + str(a_desc))
def save_action_results(self, a_desc, sn_hash, reward_val,
force_deterministic=False):
"""
Add sn_hash to possible next states and add to its RunningAve
If force_deterministic is True, force the new sn_hash to be unique
"""
# make sure that a_desc is initialized
if a_desc not in self.action_countD:
self.add_action( a_desc )
# increment action counters
self.action_countD[ a_desc ] += 1 # inc. count of a_desc calls
self.total_action_calls += 1
# make sure sn_hash dict is initialized for a_desc
if a_desc not in self.action_sn_rD:
self.action_sn_rD[ a_desc ] = {}
# snD... index=sn_hash: value=RunningAve of Reward
# save sn_hash and update reward running average for (a_desc, sn_hash)
if sn_hash not in self.action_sn_rD[ a_desc ]:
self.action_sn_rD[ a_desc ][ sn_hash ] = \
RunningAve( name= 'Reward (%s, %s, %s)'%(str(self.s_hash), str(a_desc), str(sn_hash)) )
# update the RunningAve of (a_desc, sn_hash) with current reward_val
self.action_sn_rD[ a_desc ][sn_hash].add_val( reward_val )
if force_deterministic and (len(self.action_sn_rD[ a_desc ])>1):
# remove any sn_hash other than the current input sn_hash
D = {sn_hash: self.action_sn_rD[ a_desc ][sn_hash]}
self.action_sn_rD[ a_desc ] = D
self.action_sn_rD[ a_desc ][sn_hash].set_all_attrib( 1, reward_val, reward_val, reward_val)
def get_td0_data():
if 'TD0_raveL' in dataD:
TD0_raveL = dataD['TD0_raveL']
Nruns = TD0_raveL[0].num_val
print(Nruns, ' of TD0_raveL found')
else:
TD0_raveL = []
Nruns = 0
for loop in range(Nruns, RUN_COUNT):
learn_tracker.clear()
policy, state_value = \
td0_epsilon_greedy( CW, learn_tracker=learn_tracker,
initial_Vs=0.0, # init non-terminal_set of V(s) (terminal_set=0.0)
use_list_of_start_states=False, # use list OR single start state of environment.
do_summ_print=False, show_last_change=False, fmt_V='%g', fmt_R='%g',
pcent_progress_print=0,
show_banner = False,
max_num_episodes=500, min_num_episodes=10, max_abserr=0.001,
gamma=1.0,
max_episode_steps=1000,
epsilon=EPSILON,
alpha=ALPHA)
reward_sum_per_episodeL = learn_tracker.reward_sum_per_episode()
while len(reward_sum_per_episodeL) > len(TD0_raveL):
TD0_raveL.append(RunningAve())
for R, r in zip(TD0_raveL, reward_sum_per_episodeL):
R.add_val(r)
dataD['TD0_raveL'] = TD0_raveL
save_to_pickle()
def get_estimated_rewards(self):
"""
Return a dictionary of estimated rewards for each state.
AND a dictionary of any special message
(Will be exact for deterministic environment)
"""
est_rD = {} # index=s_hash, value=float reward estimate.
msgD = {} # index=s_hash, value=any special message
# initialize all rewards to zero for all states.
for S in self.SC.iter_states():
est_rD[S.hash] = RunningAve(S.hash)
for s_hash, a_desc, T in self.TC.iter_all_transitions():
for sn_hash, t_prob, reward in T.iter_sn_hash_prob_reward():
Robj = T.get_reward_obj(sn_hash)
if Robj.reward_type == CONST:
est_rD[sn_hash].add_val(reward)
else:
msgD[sn_hash] = 'est'
# if the reward is stochastic, average 100 values
for i in range(100):
est_rD[sn_hash].add_val(Robj())
# Need to convert RunningAve objects to float
for (s_hash, RA) in est_rD.items():
est_rD[s_hash] = RA.get_ave()
#print(s_hash, RA)
return est_rD, msgD
def get_expected_sarsa_data():
if 'ExpSarsa_raveD' in dataD:
ExpSarsa_raveD = dataD['ExpSarsa_raveD']
ave_run_time = dataD['ExpSarsa_ave_run_time']
else:
ExpSarsa_raveD = {}
ave_run_time = RunningAve()
for alpha in ALPHA_LIST:
ExpSarsa_raveD[alpha] = [RunningAve(), RunningAve()]
Nruns = ExpSarsa_raveD[0.1][0].num_val
print(Nruns, ' of ExpSarsa_raveD found')
for loop in range(Nruns, RUN_COUNT):
for alpha in ALPHA_LIST:
start_time = time.time()
learn_tracker.clear()
policy, state_value = \
expected_sarsa_eps_greedy( CW, learn_tracker=learn_tracker,
initial_Qsa=0.0, # init non-terminal_set of V(s) (terminal_set=0.0)
use_list_of_start_states=False, # use list OR single start state of environment.
do_summ_print=False, show_last_change=False, fmt_Q='%g', fmt_R='%g',
pcent_progress_print=0,
show_banner = False,
max_num_episodes=1000, min_num_episodes=1000, max_abserr=0.000001,
gamma=1.0,
max_episode_steps=10000,
epsilon=EPSILON,
alpha=alpha)
reward_sum_per_episodeL = learn_tracker.reward_sum_per_episode()
ave_run_time.add_val(time.time() -
start_time) # compute average run time
ExpSarsa_raveD[alpha][0].add_val(
sum(reward_sum_per_episodeL[:100]) / 100.0)
ExpSarsa_raveD[alpha][1].add_val(
sum(reward_sum_per_episodeL) / 1000.0)
print('.', end='')
print('ExpSarsa_ave_run_time = ', ave_run_time.get_ave())
dataD['ExpSarsa_raveD'] = ExpSarsa_raveD
dataD['ExpSarsa_ave_run_time'] = ave_run_time
save_to_pickle('ExpSarsa_raveD', 'ExpSarsa_ave_run_time')
#print('true_valueD =',true_valueD)
#sys.exit()
# ----------------------------------------- generate data -------------
alphaL = [0.05*n for n in range(21)]
nstepL = [1,2,4,8, 16, 32]
nstep_walkerL = []
ave_rms_aveD = {} # index=(alpha, Nsteps), value=RunningAve
sv_collD = {} # index=(alpha, Nsteps), value=StateValueColl
# create data structures
for Nsteps in nstepL:
nstep_walkerL.append( NStepTDWalker(rw_mrp, Nsteps=Nsteps, episode_obj=episode_obj) )
for alpha in alphaL:
ave_rms_aveD[ (alpha, Nsteps) ] = RunningAve()
sv_collD[ (alpha, Nsteps) ] = StateValueColl( rw_mrp, init_val=0.0 )
# begin main loop over runs
for loop in range(AVE_OVER): # average rms curves over AVE_OVER runs
if loop%10==0:
print(loop, end='')
else:
print('.', end='')
# set state variables to 0.0
for Nsteps in nstepL :
for alpha in alphaL:
sv_collD[ (alpha, Nsteps) ].init_Vs_to_val( 0.0 )
# get the initial RMS
expected_sarsa_eps_greedy( CW, learn_tracker=learn_tracker,
initial_Qsa=0.0, # init non-terminal_set of V(s) (terminal_set=0.0)
use_list_of_start_states=False, # use list OR single start state of environment.
do_summ_print=False, show_last_change=False, fmt_Q='%g', fmt_R='%g',
show_banner = False,
pcent_progress_print=0,
max_num_episodes=500, min_num_episodes=10, max_abserr=0.001,
gamma=1.0,
max_episode_steps=1000,
epsilon=EPSILON,
alpha=ALPHA)
reward_sum_per_episodeL_es = learn_tracker.reward_sum_per_episode()
while len(reward_sum_per_episodeL_es) > len(ExpSarsa_raveL):
ExpSarsa_raveL.append(RunningAve())
for R, r in zip(ExpSarsa_raveL, reward_sum_per_episodeL_es):
R.add_val(r)
learn_tracker.clear()
policy_t, state_value_t = \
td0_epsilon_greedy( CW, learn_tracker=learn_tracker,
initial_Vs=0.0, # init non-terminal_set of V(s) (terminal_set=0.0)
use_list_of_start_states=False, # use list OR single start state of environment.
do_summ_print=False, show_last_change=False, fmt_V='%g', fmt_R='%g',
show_banner = False,
pcent_progress_print=0,
max_num_episodes=500, min_num_episodes=10, max_abserr=0.001,
gamma=1.0,
max_episode_steps=1000,
epsilon=EPSILON,
import matplotlib
import matplotlib.pyplot as plt
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.mdp_data.random_walk_mrp import get_random_walk
from introrl.agent_supt.episode_maker import make_episode
from introrl.utils.running_ave import RunningAve
rw_mrp = get_random_walk()
policy = Policy( environment=rw_mrp )
NumEpisodes = 100
mc_rms_raveL = [RunningAve(name='%i'%i) for i in range(NumEpisodes)]
td_rms_raveL = [RunningAve(name='%i'%i) for i in range(NumEpisodes)]
alpha_td = 0.1
alpha_mc = 0.02
gamma = 1.0
true_valueD = {'A':1.0/6.0, 'B':2.0/6.0, 'C':3.0/6.0, 'D':4.0/6.0, 'E':5.0/6.0}
for o_loop in range(1,101):
print('%2i'%o_loop, end=' ')
if o_loop % 20 == 0:
print()
# make 2 state value objects.
sv_td = StateValueColl( rw_mrp, init_val=0.5 )
sv_mc = StateValueColl( rw_mrp, init_val=0.5 )
def init_Vs_to_zero(self):
# initialize to 0.0 for all states, terminal and non-terminal.
for s_hash in self.environment.iter_all_states():
self.Vs_RaveD[s_hash] = RunningAve(name=s_hash)
sarsa_epsilon_greedy( CW, learn_tracker=learn_tracker,
initial_Qsa=0.0, # init non-terminal_set of V(s) (terminal_set=0.0)
use_list_of_start_states=False, # use list OR single start state of environment.
do_summ_print=False, show_last_change=False, fmt_Q='%g', fmt_R='%g',
pcent_progress_print=0,
show_banner = False,
max_num_episodes=500, min_num_episodes=10, max_abserr=0.001,
gamma=1.0,
max_episode_steps=1000,
epsilon=EPSILON,
alpha=ALPHA)
reward_sum_per_episodeL_s = learn_tracker.reward_sum_per_episode()
while len(reward_sum_per_episodeL_s) > len(Sarsa_raveL):
Sarsa_raveL.append( RunningAve() )
for R,r in zip(Sarsa_raveL, reward_sum_per_episodeL_s):
R.add_val( r )
learn_tracker.clear()
policy_q, state_value_q = \
qlearning_epsilon_greedy( CW, learn_tracker=learn_tracker,
initial_Qsa=0.0, # init non-terminal_set of V(s) (terminal_set=0.0)
use_list_of_start_states=False, # use list OR single start state of environment.
do_summ_print=False, show_last_change=False, fmt_Q='%g', fmt_R='%g',
pcent_progress_print=0,
show_banner = False,
max_num_episodes=500, min_num_episodes=10, max_abserr=0.001,
gamma=1.0,
max_episode_steps=1000,
maze_q.open_gate_R()
maze_q.close_gate_L()
# DynaQ+ episodes
while agent_qp.model.total_action_calls < 3000:
if agent_qp.model.total_action_calls >= 1000:
maze_q.open_gate_L()
maze_q.close_gate_R()
agent_qp.run_episode('Start', Nplanning_loops=PLAN_LOOPS)
cum_rew_qL = learn_tracker_q.cum_reward_per_step()
cum_rew_qpL = learn_tracker_qp.cum_reward_per_step()
while len(q_raveL) < min(3000, len(cum_rew_qL)):
q_raveL.append(RunningAve())
for i, r in enumerate(cum_rew_qL):
if i < 3000:
q_raveL[i].add_val(r)
while len(qp_raveL) < min(3000, len(cum_rew_qpL)):
qp_raveL.append(RunningAve())
for i, r in enumerate(cum_rew_qpL):
if i < 3000:
qp_raveL[i].add_val(r)
#agent_q.model.summ_print(long=True)
#sys.exit()
agent_q.action_value_coll.summ_print(fmt_Q='%.3f',
none_str='*',
show_states=True,
import sys
import matplotlib
import matplotlib.pyplot as plt
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.mdp_data.random_walk_mrp import get_random_walk
from introrl.agent_supt.episode_maker import make_episode
from introrl.utils.running_ave import RunningAve
rw_mrp = get_random_walk()
policy = Policy( environment=rw_mrp )
NumEpisodes = 100
mc_rms_raveL = [RunningAve(name='%i'%i) for i in range(NumEpisodes)]
td_rms_raveL = [RunningAve(name='%i'%i) for i in range(NumEpisodes)]
alpha = 0.1
gamma = 1.0
true_valueD = {'A':1.0/6.0, 'B':2.0/6.0, 'C':3.0/6.0, 'D':4.0/6.0, 'E':5.0/6.0}
def calc_td_error(show_values=True):
errD = {} # index=s_hash, value=Vtarget - V(s)
for s_hash in ['A','B','C','D','E']:
errD[s_hash] = 0.0
for (s_hash,sn_hash), R in td_averD.items():
errD[s_hash] += R.get_ave() + gamma*sv_td.get_Vs(sn_hash) - sv_td.get_Vs(s_hash)
fig, ax = plt.subplots()
# ---------------- set up true value data for RMS calc --------------------
true_valueD = {'C': 0.0, 'Win': 0.0, 'Lose': 0.0}
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):
d += delta
true_valueD = {'L-%i' % i: -d}
true_valueD = {'R+%i' % i: d}
# ----------------------------------------- generate TD(0) data -------------
alphaL = [0.01] + [0.05 * n for n in range(1, 21)]
ave_rms_aveL = [RunningAve(name='alpha=%g' % alpha) for alpha in alphaL]
for ialpha, alpha in enumerate(alphaL):
for loop in range(100): # average rms curves over 100 runs
sv = StateValueColl(rw_mrp, init_val=0.5)
resultL, value_snapD = td0_prediction(
policy,
sv,
all_start_states=False,
do_summ_print=False,
show_last_change=False,
show_banner=False,
pcent_progress_print=0,
alpha=alpha,
