def __init__(self,
name='Generic Environment',
s_hash_rowL=None,
row_tickL=None,
x_axis_label='',
col_tickL=None,
y_axis_label='',
colorD=None,
basic_color='',
mdp_file=None):
"""
A Basic Environment from which all others derive.
If "s_hash_rowL" is input, it will be used to calc state layout
colorD and basic_color affect the GenericLayout when saved as an image.
"""
self.name = name
self.define_statesD = {
} # index=s_hash: value=DefineStateMoves object for s_hash
self.TC = TransitionColl(name=name + ' TransitionColl')
self.default_policyD = None # may define later.
# for convenience, make TransitionColl objects available locally
self.SAC = self.TC.sa_coll # share sa_coll with TransitionColl
self.AC = self.TC.action_coll # share action_coll with TransitionColl
self.SC = self.TC.state_coll # share state_coll with TransitionColl
self.define_environment()
# can define a start state list smaller than all action states
# (to create it, call define_limited_start_state_list( state_list )
self.defined_limited_start_state_list = None
self.terminal_set, self.action_state_set = self.TC.get_terminal_set_and_action_set(
)
self.info = """A Basic Environment for solving Reinforcement Learning Problems."""
self.layout = GenericLayout(self,
s_hash_rowL=s_hash_rowL,
row_tickL=row_tickL,
x_axis_label=x_axis_label,
col_tickL=col_tickL,
y_axis_label=y_axis_label,
colorD=colorD,
basic_color=basic_color)
self.failed_mdp_file_read = False
if mdp_file is not None:
if not self.read_pickle_file(mdp_file):
print('WARNING... FAILED TO OPEN MDP FILE:', mdp_file)
print('=' * 66)
print('=' * 66)
print('=' * 66)
#sys.exit()
self.failed_mdp_file_read = True
def get_env():
env = EnvBaseline(name='Simple Six State World')
env.set_info('Simple Six State World')
actionD = {
'A': ('U', ),
'B': ('ur', 'D'),
'<C>': ('ur', 'dl'),
'D': ('ur', 'ul')
}
rewardD = {'A': -1.0, 'E': 0.5, 'F': 1.0}
for (s_hash, moveL) in actionD.items():
for a_desc in moveL:
env.add_action(s_hash, a_desc, a_prob=1.0)
def add_event(s_hash, a_desc, sn_hash):
r = rewardD.get(sn_hash, 0.0)
env.add_transition(s_hash, a_desc, sn_hash, t_prob=1.0, reward_obj=r)
add_event('A', 'U', 'B')
#add_event( 'A', 'Te', 'E' )
add_event('B', 'D', 'A')
add_event('B', 'ur', '<C>')
add_event('<C>', 'dl', 'B')
add_event('<C>', 'ur', 'D')
add_event('D', 'ur', 'F')
add_event('D', 'ul', 'E')
env.define_env_states_actions(
) # send all states and actions to environment
# --------------------
s_hash_rowL = [('*', 'E', '*', 'F'), ('*', '*', 'D', '*'),
('*', '<C>', '*', '*'), ('B', '*', '*', '*'),
('A', '*', '*', '*')]
env.layout = GenericLayout(env, s_hash_rowL=s_hash_rowL)
env.start_state_hash = '<C>'
# define default_policyD
policyD = {} # index=state_hash, value=action_desc
policyD['B'] = 1
policyD['<C>'] = 1
policyD['D'] = 1
env.default_policyD = policyD
return env
def get_random_walk():
env = EnvBaseline(name='Random Walk MRP') # GenericLayout set below
env.set_info('Random Walk MRP')
actionD = {
'A': ('L', 'R'),
'B': ('L', 'R'),
'C': ('L', 'R'),
'D': ('L', 'R'),
'E': ('L', 'R')
}
rewardD = {'Win': 1.0, 'Lose': 0.0}
for (s_hash, moveL) in actionD.items():
for a_desc in moveL:
env.add_action(s_hash, a_desc, a_prob=1.0)
def add_event(s_hash, a_desc, sn_hash):
#print('s_hash, a_desc, sn_hash',s_hash, a_desc, sn_hash)
r = rewardD.get(sn_hash, 0.0)
env.add_transition(s_hash, a_desc, sn_hash, t_prob=1.0, reward_obj=r)
mrpL = ['Lose', 'A', 'B', 'C', 'D', 'E', 'Win']
for i, ci in enumerate(mrpL[1:-1]):
add_event(ci, 'L', mrpL[i])
add_event(ci, 'R', mrpL[i + 2])
env.define_env_states_actions(
) # send all states and actions to environment
# --------------------
s_hash_rowL = [mrpL]
env.layout = GenericLayout(env, s_hash_rowL=s_hash_rowL)
env.start_state_hash = 'C'
# define default_policyD
policyD = {} # index=state_hash, value=action_desc
policyD['A'] = ('L', 'R')
policyD['B'] = ('L', 'R')
policyD['C'] = ('L', 'R')
policyD['D'] = ('L', 'R')
policyD['E'] = ('L', 'R')
env.default_policyD = policyD
return env
def define_environment(self):
for state_hash in range(1, 3):
for action_desc in [-1, 1]:
self.add_action(state_hash, action_desc,
a_prob=1.0) # a_prob will be normalized
sn = state_hash + action_desc
self.add_transition(state_hash,
action_desc,
sn,
t_prob=1.0,
reward_obj=rt)
self.define_env_states_actions(
) # send all states and actions to environment
self.start_state_hash = 12
self.layout = GenericLayout(self)
def __init__(self,
name='Basic Sim',
s_hash_rowL=None,
row_tickL=None,
col_tickL=None,
x_axis_label='',
y_axis_label='',
colorD=None,
basic_color='',
start_time=0):
"""
A Black Box Interface to a Simulation
"""
self.name = name
self.info = """A Black Box Interface to a Simulation."""
self.s_hash_rowL = s_hash_rowL
self.row_tickL = row_tickL
self.col_tickL = col_tickL
self.x_axis_label = x_axis_label
self.y_axis_label = y_axis_label
self.colorD = colorD
self.basic_color = basic_color
# state hash is (# cars at 1st site, # cars at 2nd site)
self.action_state_set = set([0, 1, 2,
3]) # a set of action state hashes
self.terminal_set = set([4]) # a set of terminal state hashes
if s_hash_rowL is None:
self.layout = None # may have a layout object for display purposes. (e.g. GenericLayout)
else:
self.layout = GenericLayout(self,
s_hash_rowL=s_hash_rowL,
row_tickL=row_tickL,
col_tickL=col_tickL,
x_axis_label=x_axis_label,
y_axis_label=y_axis_label,
colorD=colorD,
basic_color=basic_color)
self.default_policyD = None # may define later.
def define_environment(self):
# possible moves: ('U','ur','R','dr','D','dl','L','ul','Te')
actionD = {
'A': ('U', ), #'Te'),
'B': ('ur', 'D'),
'<C>': ('ur', 'dl'),
'D': ('ur', 'ul')
}
rewardD = {'A': -1.0, 'E': 0.5, 'F': 1.0}
for (s_hash, moveL) in actionD.items():
for a_desc in moveL:
self.add_action(s_hash, a_desc, a_prob=1.0)
def add_event(s_hash, a_desc, sn_hash):
r = rewardD.get(sn_hash, 0.0)
self.add_transition(s_hash,
a_desc,
sn_hash,
t_prob=1.0,
reward_obj=r)
add_event('A', 'U', 'B')
#add_event( 'A', 'Te', 'E' )
add_event('B', 'D', 'A')
add_event('B', 'ur', '<C>')
add_event('<C>', 'dl', 'B')
add_event('<C>', 'ur', 'D')
add_event('D', 'ur', 'F')
add_event('D', 'ul', 'E')
self.define_env_states_actions(
) # send all states and actions to environment
s_hash_rowL = [('*', 'E', '*', 'F'), ('*', '*', 'D', '*'),
('*', '<C>', '*', '*'), ('B', '*', '*', '*'),
('A', '*', '*', '*')]
self.layout = GenericLayout(self, s_hash_rowL=s_hash_rowL)
def get_gridworld(step_reward=-1, height=7, goal=(3,7),
windT=(0,0,0,1,1,1,2,2,1,0)):
"""
Windy Gridworld with (0,0) at lower left
width is defined by length of windT tuple.
"""
gridworld = EnvBaseline( name='Windy Kings Gridworld' ) # GenericLayout set below
gridworld.set_info( """""" )
width = len( windT )
def get_action_snext( s_hash, action):
"""returns state_next_hash"""
di = 0
dj = 0
if 'N' in action:
di = 1
elif 'S' in action:
di = -1
if 'E' in action:
dj = 1
elif 'W' in action:
dj = -1
(i,j) = s_hash
wind_di = windT[ j ]
i_next = i + di
# constrain basic move to be inside the grid
i_next = max(0, min(height-1, i_next))
i_next += wind_di # add wind to constrained move.
j_next = j + dj
# constrain next position to be inside the grid
i_next = max(0, min(height-1, i_next))
j_next = max(0, min(width-1, j_next))
state_next_hash = (i_next, j_next)
if state_next_hash == goal:
state_next_hash = 'Goal'
return state_next_hash
# define default policy
gridworld.default_policyD = {} #index=s_hash, value=list of equiprobable actions
for i in range(height):
for j in range(width):
s_hash = (i,j)
if s_hash == goal:
pass # s_hash == 'Goal'
else:
gridworld.default_policyD[ s_hash ] = ('N','S','E','W', 'NE','SE','SW','NW')
for a_desc in ['N','S','E','W', 'NE','SE','SW','NW']:
gridworld.add_action( s_hash, a_desc, a_prob=1.0 ) # a_prob will be normalized
sn_hash = get_action_snext( s_hash, a_desc )
# add each event to transitions object
gridworld.add_transition( s_hash, a_desc, sn_hash, t_prob=1.0, reward_obj=step_reward)
gridworld.define_env_states_actions() # send all states and actions to environment
# --------------------
s_hash_rowL = [] # layout rows for makeing 2D output
for i in range(height): # put (0,0) at upper left
rowL = []
for j in range(width):
s_hash = (i,j)
if s_hash == goal:
s_hash = 'Goal'
rowL.append( s_hash )
# use insert to put (0,0) at lower left, append for upper left
s_hash_rowL.insert(0, rowL )# layout rows for makeing 2D output
gridworld.layout = GenericLayout( gridworld, s_hash_rowL=s_hash_rowL,
col_tickL=windT,
x_axis_label='Upward Wind Speed' )
gridworld.start_state_hash = (3,0)
return gridworld
def get_gridworld(step_reward=0.0):
gridworld = EnvBaseline(
name='Simple Grid World') # GenericLayout set below
gridworld.set_info('Simple Grid World Example.')
actionD = {
(0, 0): ('D', 'R'),
(0, 1): ('L', 'R'),
(0, 2): ('L', 'D', 'R'),
(1, 0): ('U', 'D'),
(1, 2): ('U', 'D', 'R'),
(2, 0): ('U', 'R'),
(2, 1): ('L', 'R'),
(2, 2): ('L', 'R', 'U'),
(2, 3): ('L', 'U')
}
rewardD = {(0, 3): 1, (1, 3): -1}
for state_hash, actionL in actionD.items():
for action_desc in actionL:
gridworld.add_action(state_hash, action_desc,
a_prob=1.0) # a_prob will be normalized
a = action_desc
s = state_hash
if a == 'U':
state_next_hash = (s[0] - 1, s[1])
elif a == 'D':
state_next_hash = (s[0] + 1, s[1])
elif a == 'R':
state_next_hash = (s[0], s[1] + 1)
elif a == 'L':
state_next_hash = (s[0], s[1] - 1)
reward_val = rewardD.get(state_next_hash, step_reward)
gridworld.add_transition(state_hash,
action_desc,
state_next_hash,
t_prob=1.0,
reward_obj=reward_val)
gridworld.define_env_states_actions(
) # send all states and actions to environment
gridworld.layout = GenericLayout(
gridworld) # uses default "get_layout_row_col_of_state"
# If there is a start state, define it here.
gridworld.start_state_hash = (2, 0)
gridworld.define_limited_start_state_list([(2, 0), (2, 2)])
# define default policy (if any)
# Policy Dictionary for: GridWorld
policyD = {} # index=state_hash, value=action_desc
# Vpi shown for gamma=0.9
policyD[(0, 0)] = 'R' # Vpi=0.81
policyD[(1, 0)] = 'U' # Vpi=0.729
policyD[(0, 1)] = 'R' # Vpi=0.9
policyD[(0, 2)] = 'R' # Vpi=1.0
policyD[(1, 2)] = 'U' # Vpi=0.9
policyD[(2, 0)] = 'U' # Vpi=0.6561
policyD[(2, 2)] = 'U' # Vpi=0.81
policyD[(2, 1)] = 'R' # Vpi=0.729
policyD[(2, 3)] = 'L' # Vpi=0.729
gridworld.default_policyD = policyD
return gridworld
def get_gridworld(step_reward=0.0,
width=9,
height=6,
goal=(0, 8),
start=(2, 0),
wallL=((1, 2), (2, 2), (3, 2), (0, 7), (1, 7), (2, 7), (4,
5))):
gridworld = EnvBaseline(
name='Sutton Ex8.1 Dyna Maze') # GenericLayout set below
gridworld.set_info("""Sutton Ex8.1 Dyna Maze""")
def get_action_snext_reward(s_hash, action):
"""returns reward and state_next_hash"""
di = 0
dj = 0
reward = 0
if action == 'U':
di = -1
elif action == 'D':
di = 1
elif action == 'R':
dj = 1
elif action == 'L':
dj = -1
(i, j) = s_hash
i_next = i + di
j_next = j + dj
if j_next >= width:
j_next = j
elif j_next < 0:
j_next = j
if i_next >= height:
i_next = i
elif i_next < 0:
i_next = i
if (i_next, j_next) in wallL:
i_next, j_next = i, j
state_next_hash = (i_next, j_next)
if state_next_hash == goal:
reward = 1.0
else:
reward = 0.0
return reward, state_next_hash
# define default policy
gridworld.default_policyD = {
} #index=s_hash, value=list of equiprobable actions
for i in range(height):
for j in range(width):
s_hash = (i, j)
if s_hash != goal:
gridworld.default_policyD[s_hash] = ('U', 'D', 'R', 'L')
for a_desc in ['U', 'D', 'R', 'L']:
gridworld.add_action(
s_hash, a_desc,
a_prob=1.0) # a_prob will be normalized
reward_val, sn_hash = get_action_snext_reward(
s_hash, a_desc)
# add each event to transitions object
gridworld.add_transition(s_hash,
a_desc,
sn_hash,
t_prob=1.0,
reward_obj=reward_val)
gridworld.define_env_states_actions(
) # send all states and actions to environment
# --------------------
s_hash_rowL = [] # layout rows for makeing 2D output
for i in range(height): # put (0,0) at upper left
rowL = []
for j in range(width):
s = (i, j)
if s in wallL:
rowL.append('"Wall"')
else:
rowL.append(s)
# use insert to put (0,0) at lower left
s_hash_rowL.append(rowL) # layout rows for makeing 2D output
named_s_hashD = {start: 'Start', goal: 'Goal'}
gridworld.layout = GenericLayout(gridworld,
s_hash_rowL=s_hash_rowL,
named_s_hashD=named_s_hashD)
gridworld.start_state_hash = start
return gridworld
class EnvBaseline(object):
def __init__(self,
name='Generic Environment',
s_hash_rowL=None,
row_tickL=None,
x_axis_label='',
col_tickL=None,
y_axis_label='',
colorD=None,
basic_color='',
mdp_file=None):
"""
A Basic Environment from which all others derive.
If "s_hash_rowL" is input, it will be used to calc state layout
colorD and basic_color affect the GenericLayout when saved as an image.
"""
self.name = name
self.define_statesD = {
} # index=s_hash: value=DefineStateMoves object for s_hash
self.TC = TransitionColl(name=name + ' TransitionColl')
self.default_policyD = None # may define later.
# for convenience, make TransitionColl objects available locally
self.SAC = self.TC.sa_coll # share sa_coll with TransitionColl
self.AC = self.TC.action_coll # share action_coll with TransitionColl
self.SC = self.TC.state_coll # share state_coll with TransitionColl
self.define_environment()
# can define a start state list smaller than all action states
# (to create it, call define_limited_start_state_list( state_list )
self.defined_limited_start_state_list = None
self.terminal_set, self.action_state_set = self.TC.get_terminal_set_and_action_set(
)
self.info = """A Basic Environment for solving Reinforcement Learning Problems."""
self.layout = GenericLayout(self,
s_hash_rowL=s_hash_rowL,
row_tickL=row_tickL,
x_axis_label=x_axis_label,
col_tickL=col_tickL,
y_axis_label=y_axis_label,
colorD=colorD,
basic_color=basic_color)
self.failed_mdp_file_read = False
if mdp_file is not None:
if not self.read_pickle_file(mdp_file):
print('WARNING... FAILED TO OPEN MDP FILE:', mdp_file)
print('=' * 66)
print('=' * 66)
print('=' * 66)
#sys.exit()
self.failed_mdp_file_read = True
def get_policy_score(self,
policy=None,
start_state_hash=None,
step_limit=1000):
"""
Given a Policy object, OR policy dictionary,
apply it to the Environment and return a score
Can iterate over limited_start_state_list, or simply start at start_state_hash.
"""
if policy is None:
policy = self.SAC
if start_state_hash is None:
s_hash = self.start_state_hash
else:
s_hash = start_state_hash
r_sum = 0.0
n_steps = 0
a_desc = policy.get(s_hash, None)
#print( policy )
while (a_desc is not None) and (n_steps < step_limit):
sn_hash, reward = self.get_action_snext_reward(s_hash, a_desc)
try: # if reward is numeric, add to r_sum
r_sum += reward
except:
pass
n_steps += 1
s_hash = sn_hash
a_desc = policy.get(s_hash, None)
msg = '' # any special message(s)
return (r_sum, n_steps, msg
) # can OVERRIDE this to return a more meaningful score.
def make_pickle_filename(self, fname):
"""Make a file name ending with .mdp_pickle """
if fname is None:
fname = self.name.replace(' ', '_').replace('.',
'_') + '.mdp_pickle'
else:
fname = fname.replace(' ', '_').replace('.', '_') + '.mdp_pickle'
return fname
def save_to_pickle_file(self, fname=None): # pragma: no cover
"""Saves data to pickle file."""
#raise ValueError( 'save_to_pickle_file is BROKEN... DO NOT USE' )
fname = self.make_pickle_filename(fname)
print('Saving Environment to pickle file:', fname)
saveD = {}
saveD['name'] = self.name
saveD['define_statesD'] = self.define_statesD
saveD['info'] = self.info
saveD['layout'] = self.layout
if hasattr(self, 'start_state_hash'):
saveD['start_state_hash'] = self.start_state_hash
if hasattr(self, 'defined_limited_start_state_list'):
saveD[
'defined_limited_start_state_list'] = self.defined_limited_start_state_list
fileObject = open(fname, 'wb')
pickle.dump(saveD, fileObject,
protocol=2) # protocol=2 is python 2&3 compatible.
fileObject.close()
def read_pickle_file(self, fname=None): # pragma: no cover
"""Reads data from pickle"""
#raise ValueError( 'read_pickle_file is BROKEN... DO NOT USE' )
fname = self.make_pickle_filename(fname)
if os.path.isfile(fname):
pass # all good
elif os.path.isfile(os.path.join(mdp_path, fname)):
fname = os.path.join(mdp_path, fname)
else:
print('Pickle File NOT found:', fname)
print('mdp_path:', mdp_path)
s = '''Try running: "introrl_build_mdp" to create MDP Pickle Files.
Type: introrl_build_mdp
at the command line.'''
banner(s, banner_char='', leftMargin=0, just='center')
return False
fileObject = open(fname, 'rb')
readD = pickle.load(fileObject)
self.name = readD['name']
self.define_statesD = readD['define_statesD']
self.info = readD['info']
self.layout = readD['layout']
if 'start_state_hash' in readD:
self.start_state_hash = readD['start_state_hash']
if 'defined_limited_start_state_list' in readD:
self.defined_limited_start_state_list = readD[
'defined_limited_start_state_list']
self.define_env_states_actions(
) # use define_statesD to initialize data structures
# ----------------------
fileObject.close()
return True
def set_info(self, info):
"""Input string that describes Environment."""
self.info = info
def get_info(self):
lmax = max([len(s) for s in self.info.split('\n')])
lmax = max(16, lmax)
return '\n' + 'INFO'.center(
lmax, '_') + '\n' + self.info + '\n' + '_' * lmax + '\n'
def add_action_dict(self, actionD):
"""
iterate through dictionary of actions calling "add_action" for each one.
actionD, index=s_hash, value=list of a_desc
"""
for s_hash, aL in actionD.items():
a_prob = 1.0 / float(len(aL))
for a_desc in aL:
self.add_action(s_hash, a_desc, a_prob=a_prob)
def add_action(self, s_hash, a_desc, a_prob=1.0):
if s_hash not in self.define_statesD:
self.define_statesD[s_hash] = DefineStateMoves(s_hash)
self.define_statesD[s_hash].add_action(a_desc, a_prob)
def add_transition(self,
s_hash,
a_desc,
snext_hash,
t_prob=1.0,
reward_obj=Reward(const=0.0)):
if s_hash not in self.define_statesD:
self.define_statesD[s_hash] = DefineStateMoves(s_hash)
self.define_statesD[s_hash].add_transition(a_desc, snext_hash, t_prob,
reward_obj)
def define_env_states_actions(self):
"""
Will Set or Add prob and Reward entries to sn_probD and sn_rewardD
action_prob controls the probability of picking an action from a list of actions.
i.e. if in state s, there can be a list of (a1,p1), (a2,p2), (a3,p3), etc.
trans_prob controls the probability of picking next state from a list of next states.
i.e. if taking action a in state s, there can be a list of (sn1,p1), (sn2,p2), (sn3,p3), etc.
The Reward object is always associated with (s,a,sn), however, it can vary with probabilty
distributions of its own.
"""
for (s_hash, DSM) in self.define_statesD.items():
DSM.add_to_environment(self)
# with TC updated, recalc terminal_set
self.terminal_set, self.action_state_set = self.TC.get_terminal_set_and_action_set(
)
def define_environment(self): # pragma: no cover
"""OVERRIDE THIS in order to define the environment."""
# set up environment with calls to:
# self.add_action( s_hash, a_desc, a_prob=1.0 )
# self.add_transition( s_hash, a_desc, sn_hash, t_prob=1.0, reward_obj=Reward( const=0.0 ))
# ... OR ... Reward Object can be replaced with constant float
# self.add_transition( s_hash, a_desc, sn_hash, t_prob=1.0, reward_obj=0.0)
# self.define_env_states_actions()
# -------------------
# layout object is usually created here in child objects.
#
#self.layout = GenericLayout( self, s_hash_rowL=None )
self.start_state_hash = (0, 0) # place holder
def limited_start_state_list(self):
"""
Return a limited list of starting states.
Normally used by agents that need to discover the various
states in an environment, like epsilon-greedy.
OVERRIDE THIS to return a list of states smaller than
ALL ACTION STATES.
"""
if self.defined_limited_start_state_list is None:
return self.get_all_action_state_hashes()
else:
return self.defined_limited_start_state_list[:] # return a copy
def define_limited_start_state_list(self, state_list):
# can define a start state list smaller than all action states
# (call define_limited_start_state_list( state_list )
self.defined_limited_start_state_list = state_list
def get_start_state_hash(self):
"""Assume that the value of start_state_hash has been set... so now just return it."""
if self.start_state_hash is None: # use a random state if none provided.
return self.SC.get_random_state(
) # NOTE: this "fall-back" might return a terminal state.
return self.start_state_hash
def get_set_of_all_terminal_state_hashes(self):
"""
Return a set of terminal state hash values. OR empty set.
(No non-terminal states should be included.)
Primarily used to detect the end of an episode.
"""
# just to make sure it's "fresh", update terminal_set
self.terminal_set, self.action_state_set = self.TC.get_terminal_set_and_action_set(
)
return self.terminal_set
def get_all_action_state_hashes(self):
"""
Return a list of action state hash values. OR empty list.
(No terminal states should be included.)
"""
return [
s_hash for s_hash in self.SC.iter_state_hash()
if s_hash not in self.terminal_set
]
def get_any_action_state_hash(self):
"""
Return a action state hash at random.
"""
return random.choice(tuple(self.action_state_set))
def get_action_snext_reward(self, s_hash, a_desc):
"""Get (next state hash, float reward) by taking action "a_desc" in state "s_hash" """
sn_hash = self.TC.get_prob_weighted_next_state_hash(s_hash, a_desc)
reward = self.TC.get_reward_value(s_hash, a_desc, sn_hash)
return sn_hash, reward # (next state hash, float reward)
def get_state_legal_action_list(self, s_hash):
"""
Return a list of possible actions from this state.
Include any actions thought to be zero probability.
(OR Empty list, if there are no actions)
"""
return self.SAC.get_list_of_all_action_desc(s_hash,
incl_zero_prob=True)
def get_default_policy_desc_dict(self):
"""
If the environment has a default policy, return it as a dictionary
index=state_hash, value=action_desc
NOTE: for deterministic policy, probability of each action is 1.0
so do not need to return tuples of (action, probability)
"""
# Policy Dictionary
if self.default_policyD is None:
return {}
else:
return self.default_policyD
def get_num_states(self):
return len(self.action_state_set) + len(self.terminal_set)
def get_num_action_states(self):
return len(self.action_state_set)
def get_num_terminal_states(self):
return len(self.terminal_set)
def iter_all_action_states(self, randomize=False):
"""iterate over all action states in environment"""
if randomize:
for s_hash in random.sample(self.action_state_set,
len(self.action_state_set)):
yield s_hash # assume none in terminal_set
else:
for s_hash in self.action_state_set:
yield s_hash # assume none in terminal_set
def iter_all_terminal_states(self):
"""iterate over all terminal states in environment"""
for s_hash in self.terminal_set:
yield s_hash # assume none in action_state_set
def is_legal_state(self, s_hash):
return s_hash in self.SC.stateD
def is_terminal_state(self, s_hash):
return s_hash in self.terminal_set
def iter_all_states(self):
"""iterate over all states in environment"""
for s_hash in self.iter_all_action_states():
yield s_hash # assume none in terminal_set
for s_hash in self.iter_all_terminal_states():
yield s_hash # assume none in action_state_set
def iter_state_hash_action_desc(self):
"""Iterate over all the (s,a) pairs in the environment"""
for s_hash, a_desc in self.TC.transitionsD.keys():
yield s_hash, a_desc
def iter_action_desc_prob(self, s_hash, incl_zero_prob=False):
"""
Iterate over all (action_desc, prob) pairs.
if incl_zero_prob==True, include actions with zero probability.
"""
for (a_desc, a_prob) in self.SAC.iter_action_desc_prob(
s_hash, incl_zero_prob=incl_zero_prob):
yield a_desc, a_prob
def iter_next_state_prob_reward(self,
s_hash,
a_desc,
incl_zero_prob=False):
"""
Iterate over all (next_state_obj, prob, reward) tuples
if incl_zero_prob==True, include actions with zero probability.
"""
T = self.TC.get_transition_obj(s_hash, a_desc)
for sn_hash, t_prob, reward in T.iter_sn_hash_prob_reward():
yield sn_hash, t_prob, reward
def get_layout_row_col_of_state(self,
s_hash): # can be s_hash OR State object
"""
--> OVERRIDE THIS FOR ANY SPECIALTY LAYOUTS <--
Normally it's best to simply input "s_hash_rowL" to define layout.
return an (i,j) tuple describing the location of s_hash in the layout.
The upper left corner is (0,0) such that:
i is the index to the row in "s_hash_rowL".
"""
# in case a State object is input instead of s_hash, simply fix it
if isinstance(s_hash, State):
s_hash = State.hash
# some grid layouts can use this default (i,j)
try:
(i, j) = s_hash
i = int(i)
j = int(j)
return i, j # (row, col)
except:
index = self.SC.get_state_co_index(s_hash)
#print('for s_hash=',s_hash,' index=',index)
if index is None:
return None, None # looks bad for (row, col)
n_states = len(self.SC)
if n_states <= 16:
return divmod(index, 4) # (row, col)
else:
len_row = 1 + int(math.sqrt(n_states))
return divmod(index, len_row) # (row, col)
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 summ_print(self, long=True):
print('___ "%s" Environment Summary ___' % self.name)
#term_set, action_set = self.TC.get_terminal_set_and_action_set()
#print('Passing term_set ',term_set,' to StateColl summ_print.', type(term_set))
self.SC.summ_print(terminal_set=self.terminal_set)
self.AC.summ_print()
if long:
self.TC.summ_print()
if self.layout is not None:
self.layout_print(vname='reward',
fmt='',
show_env_states=True,
none_str='*')
def layout_print(self,
vname='reward',
fmt='',
show_env_states=True,
none_str='*'):
"""print the value "vname" formatted by the environment layout (if present). """
if self.layout is None:
print(
'...ERROR... "%s" tried to layout_print w/o a defined layout' %
self.name)
return
if show_env_states:
self.layout.s_hash_print(none_str=none_str)
msgD = {} # initialize special message dictionary to empty
if vname == 'reward':
valD, msgD = self.get_estimated_rewards(
) # index=s_hash, value=float reward estimate.
else:
valD = {} # empty if not recognized vname
x_axis_label = self.layout.x_axis_label
y_axis_label = self.layout.y_axis_label
row_tickL = self.layout.row_tickL
col_tickL = self.layout.col_tickL
rows_outL = []
for row in self.layout.s_hash_rowL:
outL = []
for s_hash in row:
if s_hash not in self.SC.stateD:
if is_literal_str(s_hash):
outL.append(s_hash[1:-1])
else:
outL.append(none_str)
else:
val = valD.get(s_hash, None)
if val is None:
outL.append(none_str)
else:
if fmt:
outL.append(fmt % val)
else:
outL.append(str(val))
if msgD.get(s_hash, ''):
outL[-1] = outL[-1] + msgD.get(s_hash, '')
rows_outL.append(outL)
if rows_outL:
print_string_rows(rows_outL,
const_col_w=True,
line_chr='_',
left_pad=' ',
y_axis_label=y_axis_label,
row_tickL=row_tickL,
col_tickL=col_tickL,
header=self.name + ' %s Summary' % vname.title(),
x_axis_label=x_axis_label,
justify='right')
def get_gridworld(
step_reward=0.0,
N_mult=1, # N_mult must be an integer.
width=9,
height=6,
goal=(0, 8),
start=(2, 0),
wallL=((1, 2), (2, 2), (3, 2), (0, 7), (1, 7), (2, 7), (4, 5))):
gridworld = EnvBaseline(
name='Sutton Ex8.4 Priority Sweep Maze') # GenericLayout set below
gridworld.set_info("""Sutton Ex8.1 Dyna Maze""")
width_big = width * N_mult
height_big = height * N_mult
gridworld.characteristic_dim = width_big + height_big * 2
# get relaxed optimal length from Zhang.
gridworld.optimal_path_len = int(14 * N_mult * 1.2) + 1
def get_action_snext_reward(s_hash, action):
"""returns reward and state_next_hash"""
di = 0
dj = 0
reward = 0
if action == 'U':
di = -1
elif action == 'D':
di = 1
elif action == 'R':
dj = 1
elif action == 'L':
dj = -1
(i, j) = s_hash
i_next = i + di
j_next = j + dj
if j_next >= width_big:
j_next = j
elif j_next < 0:
j_next = j
if i_next >= height_big:
i_next = i
elif i_next < 0:
i_next = i
if (i_next, j_next) in wall_set:
i_next, j_next = i, j
state_next_hash = (i_next, j_next)
if state_next_hash in goal_set:
reward = 1.0
else:
reward = 0.0
return reward, state_next_hash
def make_big_set(pos):
"""Take an (i,j) position, pos, and expand to new, big size in x and y"""
pos_set = set()
ip, jp = pos
ip *= N_mult
jp *= N_mult
for ixn in range(N_mult):
for jxn in range(N_mult):
pos_set.add((ip + ixn, jp + jxn))
return pos_set
# define default policy
gridworld.default_policyD = {
} #index=s_hash, value=list of equiprobable actions
# redefine start
istart, jstart = start
start = (istart * N_mult, jstart * N_mult)
# make goal set
goal_set = make_big_set(goal)
# make wall set
wall_set = set()
for wall in wallL:
wall_set.update(make_big_set(wall))
# create state hash entries
for i in range(height_big):
for j in range(width_big):
s_hash = (i, j)
if (s_hash not in wall_set) and (s_hash not in goal_set):
gridworld.default_policyD[s_hash] = ('U', 'D', 'R', 'L')
for a_desc in ['U', 'D', 'R', 'L']:
gridworld.add_action(
s_hash, a_desc,
a_prob=1.0) # a_prob will be normalized
reward_val, sn_hash = get_action_snext_reward(
s_hash, a_desc)
# add each event to transitions object
gridworld.add_transition(s_hash,
a_desc,
sn_hash,
t_prob=1.0,
reward_obj=reward_val)
gridworld.define_env_states_actions(
) # send all states and actions to environment
# --------------------
s_hash_rowL = [] # layout rows for makeing 2D output
for i in range(height_big): # put (0,0) at upper left
rowL = []
for j in range(width_big):
s = (i, j)
if s in wall_set:
rowL.append('"Wall"')
else:
rowL.append(s)
# use insert to put (0,0) at lower left
s_hash_rowL.append(rowL) # layout rows for makeing 2D output
named_s_hashD = {}
named_s_hashD[start] = 'Start'
for g in goal_set:
named_s_hashD[g] = 'Goal'
gridworld.layout = GenericLayout(gridworld,
s_hash_rowL=s_hash_rowL,
named_s_hashD=named_s_hashD)
gridworld.start_state_hash = start
return gridworld
def get_env():
env = EnvBaseline( name="Jacks Car Rental (const rtn)" ) # GenericLayout set below
simplified_str ="""Shangtong Zhang's simplified model such that
the # of cars returned in daytime becomes constant
rather than a random value from poisson distribution, which will reduce calculation time
and leave the optimal policy/value state matrix almost the same"""
env.set_info( 'Example 4.2 from Sutton & Barto 2nd Edition page 81.\n' + simplified_str )
# define all possible actions.
saL = [] # a list of (s1, s2, adesc)
s_hash_rowL = [] # layout rows for makeing 2D output
for s1 in range( MAX_CARS + 1 ): # 20 cars max
rowL = [] # row of s_hash_rowL
for s2 in range( MAX_CARS + 1 ): # 20 cars max
s_hash = (s1, s2)
rowL.append( s_hash )
for a_desc in range(-5, 6): # -5 moves 5 cars from 2nd to 1st. +5 from 1st to 2nd.
if a_desc < 0: # can only move cars if they are present
if (abs(a_desc) <= s2):
env.add_action( s_hash, a_desc, a_prob=1.0 )
saL.append( (s1, s2, a_desc) )
else:
if (a_desc <= s1): # can only move cars if they are present
env.add_action( s_hash, a_desc, a_prob=1.0 )
saL.append( (s1, s2, a_desc) )
# use insert to put (0,0) at lower left
s_hash_rowL.insert(0, rowL )# layout rows for makeing 2D output
# ------------------------------
# figure out transition probabilities and rewards
for s1 in range( MAX_CARS + 1 ):
for s2 in range( MAX_CARS + 1 ):
for a_desc in range( -5, 6 ):
get_prob_reward( s1, s2, a_desc)
# ------------------------------
print('\nStarting to define car rental transitions')
# with all the probability figured out, define all transitions
for (s1, s2, a_desc, sn_hash), t_prob in total_probD.items():
txr = sum_prob_x_rewardD[ (s1, s2, a_desc, sn_hash) ]
rval = txr / t_prob
env.add_transition( (s1,s2), a_desc, sn_hash, t_prob=t_prob, reward_obj=rval)
#if s1==10 and s2==10:
# print('for (10,10) a_desc=',a_desc,' sn_hash=',sn_hash,' t_prob=',t_prob,' rval=',rval)
print('Calling: env.define_env_states_actions')
env.define_env_states_actions() # send all states and actions to environment
print('Environment Ready.')
# If there is a start state, define it here.
env.start_state_hash = (10,10)
# define default policy (if any)
env.default_policyD = {}
# --------------------
# define layout for output
env.layout = GenericLayout( env, s_hash_rowL=s_hash_rowL,
x_axis_label='#Cars at Second Location',
y_axis_label='#Cars at First Location')
return env
def get_random_walk(Nside_states=9,
win_reward=1.0,
lose_reward=-1.0,
step_reward=0.0):
Nstates = 2 * Nside_states + 1
s = '(L%i, R%i)' % (Nside_states, Nside_states)
env = EnvBaseline(name='%i State Random Walk MRP' % Nstates +
s) # GenericLayout set below
env.set_info('%i State Random Walk MRP' % Nstates + s)
RstateL = ['R+%i' % i for i in range(1, Nside_states + 1)]
LstateL = list(reversed([s.replace('R+', 'L-') for s in RstateL]))
actionD = {}
for s in LstateL:
actionD[s] = ('L', 'R')
actionD['C'] = ('L', 'R')
for s in RstateL:
actionD[s] = ('L', 'R')
rewardD = {'Win': win_reward, 'Lose': lose_reward}
for (s_hash, moveL) in actionD.items():
for a_desc in moveL:
env.add_action(s_hash, a_desc, a_prob=1.0)
def add_event(s_hash, a_desc, sn_hash):
#print('s_hash, a_desc, sn_hash',s_hash, a_desc, sn_hash)
r = rewardD.get(sn_hash, step_reward)
env.add_transition(s_hash, a_desc, sn_hash, t_prob=1.0, reward_obj=r)
mrpL = ['Lose'] + LstateL + ['C'] + RstateL + ['Win']
for i, ci in enumerate(mrpL[1:-1]):
add_event(ci, 'L', mrpL[i])
add_event(ci, 'R', mrpL[i + 2])
env.define_env_states_actions(
) # send all states and actions to environment
# -------------------- make layout for printing ------------------
s_hash_rowL = [mrpL]
env.layout = GenericLayout(env, s_hash_rowL=s_hash_rowL)
env.start_state_hash = 'C'
# define default_policyD
policyD = {} # index=state_hash, value=action_desc
policyD['C'] = ('L', 'R')
for s in LstateL:
policyD[s] = ('L', 'R')
for s in RstateL:
policyD[s] = ('L', 'R')
env.default_policyD = policyD
return env