def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
"*** YOUR CODE HERE ***"
currentIterationCounter = 1
for state in mdp.getStates():
self.values[state] = mdp.getReward(state, 'Stop', state)
while (currentIterationCounter != self.iterations):
newValues = util.Counter()
for state in mdp.getStates():
tempValues = util.Counter()
for action in mdp.getPossibleActions(state):
for newStateAndProb in mdp.getTransitionStatesAndProbs(state, action):
newState = newStateAndProb[0]
prob = newStateAndProb[1]
tempValues[action] += prob*(mdp.getReward(state, action, newState)+self.discount*self.values[newState])
newValues[state] = tempValues[tempValues.argMax()]
currentIterationCounter += 1
for state in mdp.getStates():
self.values[state] = newValues[state]
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
for i in range(iterations):
lastValues = copy.deepcopy(self.values)
for s in mdp.getStates():
aCounter = util.Counter()
for a in mdp.getPossibleActions(s):
for s2 in mdp.getStates():
aCounter[a] += self.T(s,a,s2) * (mdp.getReward(s,a,s2) + discount*lastValues[s2])
self.values[s] = aCounter[aCounter.argMax()]
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
# Write value iteration code here
# Add all states to the dictionary and initialize their values to 0.
for state in mdp.getStates():
self.values[state] = 0
# Run the evaluation a specified number of times.
for index in range(self.iterations):
# Keep self.values static during an iteration
iterationValues = util.Counter()
for state in mdp.getStates():
QValues = util.Counter()
for action in self.mdp.getPossibleActions(state):
QValues[action] = self.computeQValueFromValues(state, action)
if len(QValues) > 0:
iterationValues[state] = QValues[QValues.sortedKeys()[0]]
# Only update self.values at the end of an iteration
self.values = iterationValues
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
# Write value iteration code here
"*** YOUR CODE HERE ***"
state = mdp.getStartState()
for i in range(0,iterations):
#print "iteration: ", i
#iterate once through all states and actions, save q-values
for state in mdp.getStates():
for action in mdp.getPossibleActions(state):
#compute qValue for each action
qValue = self.getQValue(state, action)
self.values[(state,action)] = qValue
#after all qValues are computed, iterate againt through states, save value from optimal policy. these values will be V* for next iteration
for state in mdp.getStates():
action = self.getAction(state)
self.values[state] = self.values[(state, action)]
"""
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
self.newvalues = util.Counter()
# Write value iteration code here
"*** YOUR CODE HERE ***"
iterationsRun = 0
while iterationsRun < iterations:
iterationsRun += 1
for state in mdp.getStates():
self.computeActionFromValues(state)
for state in mdp.getStates():
self.values[state] = self.newvalues[state]
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
"*** YOUR CODE HERE ***"
for s in mdp.getStates():
self.values[s] = 0
"for a in mdp.getPossibleActions(s):"
"for ac in mdp.getTransitionStatesAndProbs(s,a):"
" print ac[0]"
"print ac[1]"
"copy_value = self.values.copy()"
"for c in mdp.getStates():"
" print copy_value[c]"
i=0
"self.states = mdp.getStates()"
while i < iterations:
copy_value = self.values.copy()
for s in mdp.getStates():
if not mdp.isTerminal(s):
self.values[s] = mdp.getReward(s,'north',s) + discount * max([sum([copy_value[s1] * p for (s1,p) in mdp.getTransitionStatesAndProbs(s,a)]) for a in mdp.getPossibleActions(s)])
i = i + 1
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
# Write value iteration code here
"*** YOUR CODE HERE ***"
self.depth = 1
self.qTable = {}
self.vTable = {}
for state in mdp.getStates():
self.vTable[state] = 0
self.qTable[state] = {}
for action in mdp.getPossibleActions(state):
self.qTable[state][action] = 0
while self.depth < self.iterations + 1:
self.tempTable = {}
for state in mdp.getStates():
self.stateValue = 0
if not mdp.isTerminal(state):
self.stateValue = -9999
for action in mdp.getPossibleActions(state):
self.Qtotal = 0
for nextState,prob in mdp.getTransitionStatesAndProbs(state,action):
self.reward = mdp.getReward(state, action, nextState)
self.Qtotal += prob * (self.reward + self.discount * self.vTable[nextState])
#print "###state:",state,"Next",nextState,"reward:",self.reward,"Qtotal",self.Qtotal,"Value:",self.vTable[nextState]
self.qTable[state][action] = self.Qtotal
#print self.qTable[state][action]
self.stateValue = max(self.stateValue,self.qTable[state][action])
else:
self.tempTable[state] = 0
self.tempTable[state] = self.stateValue
self.vTable = self.tempTable
self.depth += 1
for state in mdp.getStates():
self.stateValue = -9999
for action in mdp.getPossibleActions(state):
self.Qtotal = 0
for nextState,prob in mdp.getTransitionStatesAndProbs(state,action):
self.reward = mdp.getReward(state, action, nextState)
self.Qtotal += prob * (self.reward + self.discount * self.vTable[nextState])
self.qTable[state][action] = self.Qtotal
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
self.delta = 0
while(self.iterations > 0):
# self.delta = 0
batchValues = util.Counter()
for state in mdp.getStates():
maxM = -10000
if mdp.isTerminal(state):
continue
for action in mdp.getPossibleActions(state):
statesProbs = mdp.getTransitionStatesAndProbs(state, action)
sumU = 0
Rs = 0
for stateProb in statesProbs:
# if stateProb[0] == 'TERMINAL_STATE':
# continue
sumU = sumU + self.values[stateProb[0]]*stateProb[1]
Rs = Rs + mdp.getReward(state, action, stateProb[0]) * stateProb[1]
# if sumU > maxM:
# maxM = sumU
v = Rs + sumU * discount
if (v > maxM):
maxM = v
batchValues[state] = maxM
self.values = batchValues
self.iterations = self.iterations - 1
self.policy = {}
for state in mdp.getStates():
if mdp.isTerminal(state):
self.policy[state] = None
continue
QValues = []
for action in mdp.getPossibleActions(state):
QValues.append(self.getQValue(state, action))
self.policy[state] = mdp.getPossibleActions(state)[QValues.index(max (QValues))]
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
# Write value iteration code here
"*** YOUR CODE HERE ***"
# fill every state with some action.
self.actions = dict()
for state in mdp.getStates():
stateActions = mdp.getPossibleActions(state)
if len(stateActions) > 0:
action = stateActions[0]
self.actions[state] = action
for i in xrange(iterations):
# make a copy of all the values.
# this copy will get modified in the for-loop,
# and at the end of the loop,
# the new values will become then real values.
nextValues = self.values.copy()
# for every state, and if it isn't a terminal state
# (you can't do any action on a terminal state):
for state in mdp.getStates():
if not mdp.isTerminal(state):
# get the best action.
action = self.computeActionFromValues(state)
self.actions[state] = action
# get the value for doing the currently stored action.
nextValues[state] = self.computeQValueFromValues(state, action)
# copy the new values over the old values.
self.values.update(nextValues)
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
"*** YOUR CODE HERE ***"
"""
i=1
for state in mdp.getStates():
print "state ", i, ": ", state
print "possible action: ", mdp.getPossibleActions(state)
i+=1
"""
self.policy = util.Counter()
self.nextStateValue = util.Counter()
states = mdp.getStates()
for state in mdp.getStates():
self.values[state] = 0
i=0
while i < self.iterations:
self.currentStateValue = self.values.copy()
for state in states:
actions = mdp.getPossibleActions(state)
max_qvalue = -99999999
take_action = None
for action in actions:
qvalue = self.getQValue(state, action)
if qvalue > max_qvalue:
max_qvalue = qvalue
take_action = action
if max_qvalue != -99999999:
self.values[state] = max_qvalue
self.policy[state] = take_action
i+=1
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
# Write value iteration code here
"*** YOUR CODE HERE ***"
states = mdp.getStates()
for x in range(0, iterations):
values = util.Counter()
for state in states:
action = self.getAction(state)
if action is not None:
values[state] = self.getQValue(state, action)
self.values = values
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
"*** YOUR CODE HERE ***"
import sys
states = mdp.getStates()
states.pop(0)
values = util.Counter()
for i in range(iterations):
for state in states:
maxValue = -sys.maxint
actions = mdp.getPossibleActions(state)
for action in actions:
value = self.getQValue(state, action)
if value > maxValue:
maxValue = value
values[state] = maxValue
self.values = values.copy()
def __init__(self, mdp, discount=0.9, iterations=100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter(
) # A Counter is a dict with default values as 0
"*** YOUR CODE HERE ***"
states = mdp.getStates()
for k in range(0, iterations):
for state in states:
actions = []
for action in mdp.getPossibleActions(state):
trans_prob = mdp.getTransitionStatesAndProbs(state, action)
actions.append(
sum(self.values[tp[0], k - 1] * tp[1]
for tp in trans_prob))
if actions:
max_prob = max(actions)
else:
max_prob = 0
self.values[state,
k] = mdp.getReward(state) + discount * max_prob
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter()
for i in range(iterations): # running the alg on the indicated number of iterations
y = self.values.copy() #V sub k-1
for state in mdp.getStates():
actions = util.Counter()
if mdp.isTerminal(state) == False:
for possibleActions in mdp.getPossibleActions(state):
for transitionState, prob in mdp.getTransitionStatesAndProbs(state, possibleActions):
value_iteration = prob * (mdp.getReward(state, possibleActions, transitionState) + (discount* y[transitionState]))
actions[possibleActions] += value_iteration
self.values[state] = actions[actions.argMax()]
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
# Write value iteration code here
"*** YOUR CODE HERE ***"
allStates = mdp.getStates()
for iter in range(0, iterations+1):
valuesIter = self.calcValue(allStates, iter)
self.values = valuesIter.copy()
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
states = mdp.getStates()
for i in range(0,self.iterations):
V = util.Counter()
for state in states:
action = self.computeActionFromValues(state)
if action is None:
V[state] = 0
else:
V[state] = self.computeQValueFromValues(state,action)
self.values = V
def __init__(self, mdp, discount=0.9, iterations=100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
# Write value iteration code here
"*** YOUR CODE HERE ***"
# Value iteration works based on the Belman equations
for _ in range(self.iterations):
values = self.values.copy()
for state in mdp.getStates():
actions = self.mdp.getPossibleActions(state)
if not actions:
values[state] = 0
else:
values[state] = max(
self.computeQValueFromValues(state, action)
for action in actions)
self.values = values
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
self.new_values = util.Counter() # A Counter is a dict with default 0
self.my_states = mdp.getStates()
#for s in mdp.getStates():
# print(s)
for i in range(0, iterations):
self.do_one_iteration(mdp)
self.values = self.new_values
print("Finished my iterations ")
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
self.startState = (0, 0)
# Write value iteration code here
"*** YOUR CODE HERE ***"
for _ in range(0, iterations):
currentValues = {}
for state in mdp.getStates():
bestAction = self.computeActionFromValues(state)
bestActionValue = self.computeQValueFromValues(state, bestAction)
currentValues[state] = bestActionValue
self.values = currentValues
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
"*** YOUR CODE HERE ***"
for n in range(iterations):
V = self.values.copy()
for s in mdp.getStates():
action_values = []
for a in mdp.getPossibleActions(s):
action_value = 0
for s_, P in mdp.getTransitionStatesAndProbs(s, a):
action_value += P * (mdp.getReward(s, a, s_) + discount * V[s_])
action_values.append(action_value)
self.values[s] = max(action_values or [0])
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
# Write value iteration code here
"*** YOUR CODE HERE ***"
currentIteration = util.Counter()
for i in range(iterations):
for state in mdp.getStates():
bestScore = float('-inf')
for action in mdp.getPossibleActions(state):
qVal = self.computeQValueFromValues(state, action)
if qVal > bestScore:
bestScore = qVal
if bestScore != float('-inf'):
currentIteration[state] = bestScore
for state in self.values:
self.values[state] = currentIteration[state]
def __init__(self, mdp, discount=0.9, iterations=100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
# Write value iteration code here
"*** YOUR CODE HERE ***"
t = mdp.getTransitionStatesAndProbs
r = mdp.getReward
for i in range(iterations):
temp = self.values.copy()
for s in mdp.getStates():
if len(mdp.getPossibleActions(s)) != 0:
self.values[s] = max([
sum([
p * (r(s, a, nexts) + discount * temp[nexts])
for nexts, p in t(s, a)
]) for a in mdp.getPossibleActions(s)
])
else:
self.values[s] = temp[s]
def __init__(self, mdp, discount=0.9, iterations=100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
T = mdp.getTransitionStatesAndProbs
R = mdp.getReward
Act = mdp.getPossibleActions
self.mdp = mdp
self.discount = discount
self.iterations = iterations
disc = discount
values = util.Counter()
states = mdp.getStates()
for _ in range(iterations):
oldval = values.copy()
for s in states:
curr = [
sum(((el[1] * (R(s, a, el[0]) + disc * oldval[el[0]])
for el in T(s, a)))) for a in Act(s)
]
values[s] = max(curr) if len(curr) else 0
self.values = values
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is action dict with default 0
"*** YOUR CODE HERE ***"
for i in range(iterations):
self.prevBatch = self.values.copy()
for state in mdp.getStates():
qValues = util.Counter()
for action in mdp.getPossibleActions(state):
for (statePrime, tValue) in mdp.getTransitionStatesAndProbs(state, action):
qValues[action] += tValue * (mdp.getReward(state, action, statePrime) + self.discount * self.prevBatch[statePrime])
self.values[state] = qValues[qValues.argMax()]
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
# Write value iteration code here
"*** YOUR CODE HERE ***"
states = mdp.getStates()
for k in range(iterations):
newValues = {}
for state in states:
actions = mdp.getPossibleActions(state)
v = util.Counter()
for action in actions:
v[action] = self.computeQValueFromValues(state, action)
newValues[state] = v[v.argMax()]
self.values = newValues
def __init__(self, mdp, discount=0.9, iterations=100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
# Write value iteration code here
"*** YOUR CODE HERE ***"
states = mdp.getStates()
temp = util.Counter()
for i in range(iterations):
for s in states:
actions = mdp.getPossibleActions(s)
v = map(lambda x: self.computeQValueFromValues(s, x), actions)
if len(v) > 0:
temp[s] = max(v)
self.values = temp.copy()
def __init__(self, mdp, discount=0.9, iterations=100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
# Write value iteration code here
for _ in range(iterations): # do value iteration for this many times
nextValues = util.Counter(
) # init values_k+1 but still use self.values as values_k
for state in mdp.getStates(): # for every state
action = self.computeActionFromValues(
state) # calculate best action
nextValues[state] = self.computeQValueFromValues(
state, action) # set value_k+1 as val for that action
self.values = nextValues # set values_k+1 as values_k for next iteration
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
"*** YOUR CODE HERE ***"
for times in range(iterations):
V = self.values.copy()
for state in mdp.getStates():
action_values = util.Counter()
for action in mdp.getPossibleActions(state):
for trans_state, prob in mdp.getTransitionStatesAndProbs(state, action):
action_values[action] += prob * (mdp.getReward( state, action, trans_state) + discount * V[trans_state])
self.values[state] = action_values[action_values.argMax()]
def __init__(self, mdp, discount=0.9, iterations=100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
# Write value iteration code here
"*** YOUR CODE HERE ***"
for i in range(iterations):
next_v = util.Counter()
for state in mdp.getStates():
if mdp.isTerminal(state): v = 0
else: v = float("-inf")
for a in mdp.getPossibleActions(state):
q = self.computeQValueFromValues(state, a)
if (q > v): v = q
next_v[state] = v
self.values = next_v
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
self.tmpValues = util.Counter();
iterationsCompleted = 0
startState = mdp.getStartState();
while (iterationsCompleted < iterations):
for state in mdp.getStates():
self.computeValue(mdp,state,discount)
for key in self.tmpValues:
self.values[key] = self.tmpValues[key]
iterationsCompleted += 1
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
self.ValuesDup = util.Counter()
# Write value iteration code here
"*** YOUR CODE HERE ***"
iterations = self.iterations
while(iterations >0):
for astate in mdp.getStates():
if mdp.isTerminal(astate)==0:
QVallist=[]
for action in mdp.getPossibleActions(astate):
QVallist += [self.computeQValueFromValues(astate, action)]
self.values[astate]=max(QVallist)
for states,value in self.values.items():
self.ValuesDup[states] = self.values[states]
iterations+=-1
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
mdpStates = mdp.getStates()
for iteration in xrange(iterations):
newValues = util.Counter()
for state in mdpStates:
if self.mdp.isTerminal(state):
continue
actionValues = -sys.maxint - 1
for action in mdp.getPossibleActions(state):
sum = 0
for transitionState, prob in mdp.getTransitionStatesAndProbs(state, action):
sum += prob*(mdp.getReward(state, action, transitionState) + discount * self.values[transitionState])
if sum > actionValues:
actionValues = sum
newValues[state] = actionValues
self.values = newValues
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
# Write value iteration code here
for _ in range(self.iterations):
nextValues = util.Counter()
for state in mdp.getStates():
actions = mdp.getPossibleActions(state)
newValue = max(self.computeQValueFromValues(state, action) for action in actions) if len(actions) else 0
nextValues[state] = newValue
self.values = nextValues
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
"*** YOUR CODE HERE ***"
for i in range(iterations):
nextValues = util.Counter()
for state in mdp.getStates():
rewardsPossible = util.Counter()
for action in mdp.getPossibleActions(state):
nextPossible = size(mdp.getTransitionStatesAndProbs(state, action))[1]
newRewards = util.Counter()
for tmpState in range(nextPossible):
nextState = mdp.getTransitionStatesAndProbs(state, action)[tmpState][0]
prob = mdp.getTransitionStatesAndProbs(state, action)[tmpState][1]
rewards = mdp.getReward(state, action, tmpState)
newRewards[tmpState] = prob * (rewards + self.discount * self.values[nextState])
rewardsPossible[action] = newRewards.totalCount()
nextValues[state] = rewardsPossible[rewardsPossible.argMax()]
self.values = nextValues
def __init__(self, mdp, discount=0.9, iterations=100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
self.previousValues = util.Counter()
# Write value iteration code here
"*** YOUR CODE HERE ***"
gameStates = mdp.getStates()
for i in xrange(self.iterations):
for state in gameStates:
totalRewardPerAction = util.Counter()
possibeActions = mdp.getPossibleActions(state)
for action in possibeActions:
totalRewardPerAction[action] = self.getTotalReward(
state, action)
self.values[state] = totalRewardPerAction[
totalRewardPerAction.argMax()]
for stateTemp in gameStates:
self.previousValues[stateTemp] = self.values[stateTemp]
def __init__(self, mdp, discountRate = 0.9, iters = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
"""
self.mdp = mdp
self.discountRate = discountRate
self.iters = iters
self.values = util.Counter() # A Counter is a dict with default 0
"""Description:
V_k+1(s) := max_a sum_(s') T(s,a,s')[R(s,a,s') + gV_k(s')]
"""
""" YOUR CODE HERE """
# We want to start at V_0(s) and work our way up.
# Currently, all values default to 0 for every state.
allStates = mdp.getStates()
for _ in range(iters):
tempValues = self.values.copy()
for state in allStates:
if self.mdp.isTerminal(state):
continue
actions = self.mdp.getPossibleActions(state)
maxVal = max(self.getQValue(state, action) for action in actions)
tempValues[state] = maxVal
self.values = tempValues.copy()
""" END CODE """
def __init__(self, mdp, discount=0.9, iterations=100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
# Write value iteration code here
"*** YOUR CODE HERE ***"
for i in range(iterations):
values = self.values.copy() #vrednosti iz prethodne iteracije
for state in mdp.getStates():
actionsCounter = util.Counter()
for action in mdp.getPossibleActions(state):
for nextState, prob in mdp.getTransitionStatesAndProbs(
state, action):
actionsCounter[action] += prob * (
mdp.getReward(state, action, nextState) +
discount * values[nextState])
self.values[state] = actionsCounter[actionsCounter.argMax()]
def __init__(self, mdp, discount=0.9, iterations=100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
# Write value iteration code here
"*** YOUR CODE HERE ***"
self.old_values = self.values.copy()
for i in range(iterations):
for state in mdp.getStates():
if not mdp.isTerminal(state):
action_value = -float('inf')
for action in mdp.getPossibleActions(state):
q_value = self.computeQValueFromValues(state, action)
action_value = max(action_value, q_value)
self.values[state] = action_value
self.old_values = self.values.copy()
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
self.qvalues = util.Counter()
self.bestact = util.Counter()
"*** YOUR CODE HERE ***"
states = mdp.getStates()
for i in range(iterations):
v = util.Counter()
for state in states:
if mdp.isTerminal(state):
continue
value = {action: sum(prob * (mdp.getReward(state,action,next_state) + discount*self.values[next_state])
for next_state, prob in mdp.getTransitionStatesAndProbs(state, action))
for action in mdp.getPossibleActions(state)}
self.bestact[state] = max(value, key=value.get)
v[state] = value[self.bestact[state]]
for action in value.keys():
self.qvalues[state,action] = value[action]
self.values = v.copy()
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
# Write value iteration code here
"*** YOUR CODE HERE ***"
# V*(s) = max sum T(s, a, s')[R(s, a, s') + gV*(s')] = qvalue
states = mdp.getStates()
for i in range(self.iterations):
newValues = util.Counter()
for state in states:
maximum = float("-inf")
actions = self.mdp.getPossibleActions(state)
for action in actions:
temp = self.computeQValueFromValues(state, action)
if temp > maximum:
maximum = temp
newValues[state] = maximum
self.values = newValues
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
# Write value iteration code here
"*** YOUR CODE HERE ***"
#while still iterations
#for each state
#for action in each state
#get Q(state,action)
#store largest (state,action) in Counter
for i in range(self.iterations):
newValues = self.values.copy() #WTF WHY THIS TOOK HOURS
for state in mdp.getStates():
v = [float("-inf")]
if not mdp.isTerminal(state):
for action in mdp.getPossibleActions(state):
v += [self.computeQValueFromValues(state,action)]
newValues[state] = max(v)
self.values = newValues
def __init__(self, mdp, discount=0.9, iterations=100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
"*** YOUR CODE HERE ***"
for i in range(iterations):
newValues = util.Counter()
for state in mdp.getStates():
values = []
for action in mdp.getPossibleActions(state):
qvalue = sum([
(discount * self.values[newState] +
mdp.getReward(state, action, newState)) * prob
for newState, prob in mdp.getTransitionStatesAndProbs(
state, action)
])
values.append(qvalue)
if len(values) > 0:
newValues[state] = max(values)
for state in self.values:
self.values[state] = newValues[state]
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
# Write value iteration code here
"*** YOUR CODE HERE ***"
for i in range(iterations):
valuesNew = util.Counter()
for state in mdp.getStates():
maxVal = -1
if not mdp.isTerminal(state):
vals = util.Counter()
for possact in mdp.getPossibleActions(state):
#value = self.computeQValueFromValues(state, possact)
#if value > maxVal:
# maxVal = value
vals[possact] = self.computeQValueFromValues(state, possact)
#valuesNew[state] = maxVal
valuesNew[state] = max(vals.values())
for st2 in valuesNew:
self.values[st2] = valuesNew[st2]
def __init__(self, mdp, discount=0.9, iterations=100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
for i in range(iterations):
newValues = self.values.copy()
for state in mdp.getStates():
if self.mdp.isTerminal(state):
continue
possible_actions = mdp.getPossibleActions(state)
decision = -2500.0
for action in possible_actions:
if (self.getQValue(state, action) > decision):
decision = self.getQValue(state, action)
newValues[state] = decision
self.values = newValues
"*** YOUR CODE HERE ***"
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
"*** YOUR CODE HERE ***"
for time in range(iterations):
values = util.Counter()
for state in mdp.getStates():
if mdp.isTerminal(state):
values[state] = 0
else:
maxValue = -INF
for action in mdp.getPossibleActions(state):
maxValue = max(maxValue, self.getQValue(state, action))
values[state] = maxValue
self.values = values
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
# Write value iteration code here
"*** YOUR CODE HERE ***"
#print mdp.getStates()
for i in range(self.iterations):
newValues = util.Counter()
for s in mdp.getStates():
options = [ self.computeQValueFromValues(s, a) for a in mdp.getPossibleActions(s)]
#if i == 2 and s == (2,2):
# for a in self.mdp.getPossibleActions(s): self.slowCompute(s,a)
if len(options):
newValues[s] = max(options)
self.values = newValues
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
for i in range(iterations):
nextValues = util.Counter()
for state in mdp.getStates():
if mdp.isTerminal(state): continue
first = True
for action in mdp.getPossibleActions(state):
qValue = 0
for (nextState, prob) in mdp.getTransitionStatesAndProbs(state, action):
reward = mdp.getReward(state, action, nextState)
qValue += prob * (reward + discount*self.values[nextState])
if first:
maxQValue = qValue
first = False
elif qValue > maxQValue:
maxQValue = qValue
nextValues[state] = maxQValue
self.values = nextValues
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
"*** YOUR CODE HERE ***"
# Init : Not required
# Value iteration
for i in range(iterations):
old_values = self.values.copy()
for state in mdp.getStates():
value_state_action = []
for action in mdp.getPossibleActions(state):
val = 0
transition = mdp.getTransitionStatesAndProbs(state,action)
for sstate,prob_s_a_ss in transition:
val += prob_s_a_ss*(mdp.getReward(state,action,sstate) + discount*old_values[sstate])
value_state_action.append(val)
if value_state_action : self.values[state] = max(value_state_action)
def __init__(self, mdp, discount = 0.9, iterations = 100): """ Your value iteration agent should take an mdp on construction, run the indicated number of iterations and then act according to the resulting policy. Some useful mdp methods you will use: mdp.getStates() mdp.getPossibleActions(state) mdp.getTransitionStatesAndProbs(state, action) mdp.getReward(state, action, nextState) mdp.isTerminal(state) """ self.mdp = mdp self.discount = discount self.iterations = iterations self.values = util.Counter() # A Counter is a dict with default 0 # Write value iteration code here "*** YOUR CODE HERE ***" for i in range(self.iterations): # Will range through all iterations to determine each Vk(s) counter = util.Counter() for state in mdp.getStates(): # Range through all possible states if self.mdp.isTerminal(state): # This means the state value cannot change in the specific iteration counter[state] = 0 else: actions = self.mdp.getPossibleActions(state) # Get all possible actions available at a given state values = [self.computeQValueFromValues(state, action) for action in actions] # Compute the values of each action counter[state] = max(values) # Choose the best possible value (and therefore action) accordingly self.values = counter # Change the initial counter to the new one determined above
def __init__(self, mdp, discount=0.9, iterations=100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
# Write value iteration code here
"*** YOUR CODE HERE ***"
self.optimalActionInState = collections.defaultdict(None)
for k in range(iterations):
lastValues = self.values.copy()
for state in mdp.getStates():
if self.mdp.isTerminal(state):
continue
maxValue = float("-inf") if mdp.getPossibleActions(state) else 0
for action in mdp.getPossibleActions(state):
theSum = 0
for nextState, prob in self.mdp.getTransitionStatesAndProbs(state, action):
R = self.mdp.getReward(state, action, nextState)
theSum += prob * (R + self.discount * lastValues[nextState])
maxValue = max(maxValue,theSum)
self.values[state] = maxValue
def __init__(self, mdp, discount=0.9, iterations=100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
# Write value iteration code here
for k in range(1, iterations + 1):
values = copy(self.values)
for state in mdp.getStates():
actions = mdp.getPossibleActions(state)
if len(actions) > 0:
self.values[state] = max([
sum([
prob * (mdp.getReward(state, action, next) +
discount * values[next])
for next, prob in mdp.getTransitionStatesAndProbs(
state, action)
]) for action in mdp.getPossibleActions(state)
])
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
"*** YOUR CODE HERE ***"
# OUR CODE HERE
#Note: I think we should use the util.Counter thing?
for times in range(0, iterations):
#values from previous iteration so we don't update over them while iterating
prevVals = self.values.copy()
#iterate through all states
for state in mdp.getStates():
#will store the action-value for the iteration
value = util.Counter()
for action in mdp.getPossibleActions(state):
for transitionState, probability in mdp.getTransitionStatesAndProbs(state, action):
#expected value, probability * reward for the state with the discount * reward
value[action] += probability * (mdp.getReward( state, action, transitionState) + discount * prevVals[transitionState])
#update the values to the new value from the iteration
#the .argMax() function returns the one with the largest value
self.values[state] = value[value.argMax()]
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
"*** YOUR CODE HERE ***"
#get the mdp states
# self.values(state) is already initialized to 0 for all states
newVals = util.Counter()
for i in range(self.iterations):
for state in mdp.getStates():
newVals[state] = mypy.newVal(self.mdp,state,self.discount,self.values)
self.values = newVals.copy()
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
# Write value iteration code here
"*** YOUR CODE HERE ***"
#states = mdp.getStates()
#values = {state: 0 for state in states}
for i in range(iterations):
previous = self.values.copy()
for state in mdp.getStates():
possibleActions = mdp.getPossibleActions(state)
if len(possibleActions) == 0: continue
results = []
for action in possibleActions:
total = 0
for (nextState, prob) in mdp.getTransitionStatesAndProbs(state,action):
total += (prob * previous[nextState])
results.append(total)
self.values[state] = mdp.getReward(state) + (discount * max(results))
def __init__(self, mdp, discount=0.9, iterations=100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
# Write value iteration code here
"*** YOUR CODE HERE ***"
for i in range(0, iterations):
valuesCopy = self.values.copy()
for state in mdp.getStates():
maximumScore = -10000
for action in mdp.getPossibleActions(state):
currQVal = self.computeQValueFromValues(state, action)
maximumScore = max(maximumScore, currQVal)
if mdp.isTerminal(state):
valuesCopy[state] = 0
else:
valuesCopy[state] = maximumScore
self.values = valuesCopy
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
helper_vector = util.Counter() # Copy of vectors to be used for batch updating
for i in range(self.iterations):
for state in mdp.getStates():
if mdp.isTerminal(state):
continue
if mdp.getPossibleActions(state):
helper_vector[state] = sum([transition[1]*(mdp.getReward(state,mdp.getPossibleActions(state)[0],transition[0])+self.discount*self.values[transition[0]])
for transition in mdp.getTransitionStatesAndProbs(state, mdp.getPossibleActions(state)[0])] )
for action in mdp.getPossibleActions(state):
helper_vector[state] = max(helper_vector[state],sum([ transition[1]*(mdp.getReward(state, action, transition[0])+self.discount*self.values[transition[0]])
for transition in mdp.getTransitionStatesAndProbs(state, action)] ))
for state in helper_vector:
self.values[state] = helper_vector[state]
def __init__(self, mdp, discount = 0.9, iterations = 100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
# Write value iteration code here
"*** YOUR CODE HERE ***"
self.vks = util.Counter()
for i in range(0,iterations):
self.vks = self.values.copy()
st = mdp.getStates()
for s in st:
a = mdp.getPossibleActions(s)
qvals = util.Counter()
for action in a:
qvals[action] = 0
stp = self.mdp.getTransitionStatesAndProbs(s,action)
for ss, prob in stp:
qvals[action] = qvals[action] + prob*(self.mdp.getReward(s,action,ss) + self.discount*(self.vks[ss]))
self.values[s] = qvals[qvals.argMax()]
def __init__(self, mdp, discount=0.9, iterations=100):
"""
Your value iteration agent should take an mdp on
construction, run the indicated number of iterations
and then act according to the resulting policy.
b
Some useful mdp methods you will use:
mdp.getStates()
mdp.getPossibleActions(state)
mdp.getTransitionStatesAndProbs(state, action)
mdp.getReward(state, action, nextState)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
"*** YOUR CODE HERE ***"
for i in xrange(0, iterations - 1):
for state in mdp.getStates():
q_values = []
for action in mdp.getPossibleActions(state):
q = self.getQValue(state, action)
q_values.append(q)
self.values[state] = self.values[state] if len(
q_values) == 0 else max(q_values)
def runValueIteration(self):
"*** YOUR CODE HERE ***"
mdp = self.mdp
state = mdp.getStates()
predecessor_list = {}
discount = self.discount
iterations = self.iterations
theta = self.theta
for temp_state in state:
predecessor_list[temp_state] = self.getpredecessor(temp_state)
pq = util.PriorityQueue()
for temp_state in state:
if not mdp.isTerminal(temp_state):
pq.push(temp_state, -self.find_difference(temp_state))
for i in range(iterations):
if pq.isEmpty():
return
cur_state = pq.pop()
if not mdp.isTerminal(cur_state):
action = self.getAction(cur_state)
self.values[cur_state] = self.getQValue(cur_state, action)
for pre in predecessor_list[cur_state]:
diff_pre = self.find_difference(pre)
if diff_pre > theta:
pq.update(pre, -diff_pre)
