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 ***"
count = 0
for j in range(self.iterations):
newVal = self.values.copy()
for s in mdp.getStates():
k = [float("-inf")]
if mdp.isTerminal(s): #checking if mdp is a terminal state
count = count + 1
elif not mdp.isTerminal(s):
for action in mdp.getPossibleActions(s): #Getting Q state
k += [self.computeQValueFromValues(s, action)
] #computing Q values
newVal[s] = max(k)
self.values = newVal
def runValueIteration(self):
"*** YOUR CODE HERE ***"
mdp = self.mdp
predecessors = {state: set() for state in mdp.getStates()}
for state in mdp.getStates():
for action in mdp.getPossibleActions(state):
for nextState, prob in mdp.getTransitionStatesAndProbs(
state, action):
if prob > 0:
predecessors[nextState].add(state)
queue = util.PriorityQueue()
for state in mdp.getStates():
if not mdp.isTerminal(state):
diff = abs(self.values[state] - max([
self.getQValue(state, action)
for action in mdp.getPossibleActions(state)
]))
queue.update(state, -diff)
for i in range(self.iterations):
if queue.isEmpty():
break
state = queue.pop()
if not mdp.isTerminal(state):
self.values[state] = max([
self.getQValue(state, action)
for action in mdp.getPossibleActions(state)
])
for pred in predecessors[state]:
diff = abs(self.values[pred] - max([
self.getQValue(pred, action)
for action in mdp.getPossibleActions(pred)
]))
if diff > self.theta:
queue.update(pred, -diff)
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)
def runValueIteration(self):
"*** YOUR CODE HERE ***"
mdp = self.mdp
values = self.values
discount = self.discount
predecessors = {}
for state in mdp.getStates():
preList = []
for preState in mdp.getStates():
for action in mdp.getPossibleActions(preState):
if state in [
pair[0]
for pair in mdp.getTransitionStatesAndProbs(
preState, action) if pair[1] > 0
]:
preList.append(preState)
break
predecessors[state] = preList
queue = util.PriorityQueue()
for s in mdp.getStates():
if not mdp.isTerminal(s):
actions = mdp.getPossibleActions(s)
realValue = max(
sum(prob * (mdp.getReward(s, action, nextState) +
(discount * values[nextState]))
for (nextState, prob
) in mdp.getTransitionStatesAndProbs(s, action))
for action in actions)
diff = abs(realValue - values[s])
queue.push(s, 0 - diff)
for _ in range(self.iterations):
if queue.isEmpty():
return
s = queue.pop()
if not mdp.isTerminal(s):
actions = mdp.getPossibleActions(s)
values[s] = max(
sum(prob * (mdp.getReward(s, action, nextState) +
(discount * values[nextState]))
for (nextState, prob
) in mdp.getTransitionStatesAndProbs(s, action))
for action in actions)
for p in predecessors[s]:
actions = mdp.getPossibleActions(p)
realValue = max(
sum(prob * (mdp.getReward(p, action, nextState) +
(discount * values[nextState]))
for (nextState, prob
) in mdp.getTransitionStatesAndProbs(p, action))
for action in actions)
diff = abs(realValue - values[p])
if diff > self.theta:
queue.update(p, 0 - diff)
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 get_average_reward(agent, mdp):
current_state = mdp.getStartState()
iterations = 0
total = 0
while not mdp.isTerminal(current_state) and iterations < 1000:
total += agent.getValue(current_state)
action = agent.getPolicy(current_state)
next_states = mdp.getTransitionStatesAndProbs(current_state, action)
current_state = next_states[0][0]
iterations += 1
if mdp.isTerminal(current_state):
return total / iterations
return 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 ***"
for ite in range(iterations): ## repeat
stateValues=util.Counter() ## initial all values are 0
for state in mdp.getStates():
if mdp.isTerminal(state)==False: ##if not a terminal state
## maximizes the expected utility of each state
stateValues[state] = max([self.computeQValueFromValues(state, action) for action in mdp.getPossibleActions(state)])
self.values=stateValues
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
states = mdp.getStates()
for _ in range(0, self.iterations):
# create temporary values for for k iteration update
tempValues = util.Counter()
# loop over states to find max q value
for state in states:
if not mdp.isTerminal(state):
actions = mdp.getPossibleActions(state)
tempValues[state] = max(
[self.getQValue(state, action) for action in actions])
# set values to k iteration update values
self.values = tempValues
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 xrange(iterations):
# do batch processing
vals = copy.deepcopy(self.values)
for s in mdp.getStates():
if mdp.isTerminal(s):
vals[s] = 0
else:
vals[s] = self.computeQValueFromValues(s, self.computeActionFromValues(s))
self.values = vals
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 i in range(iterations):
# Make a new dictionary to memorize new values
newValues = util.Counter()
# For every state...
for curState in states:
if mdp.isTerminal(curState): continue
action = self.getAction(curState)
# Compute new Q value
newValues[curState] = self.computeQValueFromValues(
curState, action)
# Memorize new values
self.values = newValues.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)
"""
"*** YOUR CODE HERE ***"
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # value of each state; a Counter is a dict with default 0
# run for desired number of iterations
for i in xrange(iterations):
new_values = self.values.copy()
for s in mdp.getStates():
if not mdp.isTerminal(s):
# the commented code works as well
#curr_best = float("-inf")
#for a in mdp.getPossibleActions(s):
#temp_value = sum([p * (mdp.getReward(s, a, s2) + discount*prev[s2]) for s2, p in mdp.getTransitionStatesAndProbs(s, a)])
# if temp_value > curr_best:
# curr_best = temp_value
#self.values[s] = curr_best
new_values[s] = max([self.getQValue(s, a) for a in mdp.getPossibleActions(s)])
self.values = new_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
v = 0
# Write value iteration code here
"*** YOUR CODE HERE ***"
for k in range(iterations):
tmp = util.Counter()
for s in mdp.getStates():
if mdp.isTerminal(s):
tmp[s] = 0
else:
maior = -9999
for a in mdp.getPossibleActions(s):
qv = self.computeQValueFromValues(s, a)
if qv > maior:
maior = qv
tmp[s] = maior
self.values = tmp
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.oldValues = self.values.copy()
states = mdp.getStates()
for i in range(iterations):
for state in states:
actions = mdp.getPossibleActions(state)
if not mdp.isTerminal(state):
actionValue = -float('inf')
for action in actions:
qValue = self.computeQValueFromValues(state, action)
actionValue = max(actionValue, qValue)
self.values[state] = actionValue
self.oldValues = 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)
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 ***"
# Outer loop for iteration time
# Define the value for each of state taking pi action
for i in range(0, iterations):
newDict = util.Counter() # Create another dictionary to prevent mutation
for state in mdp.getStates():
if mdp.isTerminal(state):
newDict[state] = 0 # Set zero when it's in terminal node
else:
newDict[state] = max([ self.getQValue(state, action) for action in mdp.getPossibleActions(state)])
self.values = newDict # Update the raw 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.
"""
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 ***"
#mdp este procesul de decizii, care este cunoscut (spre deosebire de q learning,
#unde nu este cunoscut modelul
for iteration in range(iterations): #se itereaza de iterations ori
currentStepValues = self.values.copy() #valorile Q ale pasului curent, deci mereu in pasul k + 1 se porneste cu valorile din pasul k
for nextState in mdp.getStates(): #toata lista de stari posibile, in cazul nostru toate patratele de pe harta
if mdp.isTerminal(nextState):#starea terminala nu isi mai schimba valoarea Q
continue
currentStepValues[nextState] = self.computeQValueFromValues(nextState, self.computeActionFromValues(nextState))#update la valori, defapt singura operatie dintr-o iteratie
#se recalculeaza valoarea q a acestei stari in functie de valoarea precedenta
#print("VALUE IS: ") # sa vedem frumos ce se intampla
#print(currentStepValues[nextState])
self.values = currentStepValues.copy()#salvam valorile noi
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)
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.oldValues = self.values.copy()
states = self.mdp.getStates()
for iteration in range(iterations):
for state in states:
if not mdp.isTerminal(state):
action = self.computeActionFromValues(state)
value = self.computeQValueFromValues(state, action)
self.values[state] = value
self.oldValues = 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)
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 ***"
stateValue = dict()
for _ in range(self.iterations):
maxVal = 0
for state in self.mdp.getStates():
if not mdp.isTerminal(state): # Check if the given state is terminal or not
actions = list() # List to store possible actions available
for action in self.mdp.getPossibleActions(state):
actions.append(self.getQValue(state, action)) # Get Q value for each of the possible states with given actions
if actions:
maxVal = max(actions) # Store the max actions value in order to calculate optimal path
stateValue[state] = maxVal
self.values = stateValue.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
#print self.values['test']
#self.values['test'] +=1
for x in range(0,iterations):
statevalues = util.Counter()
for state in mdp.getStates():
if mdp.isTerminal(state):
statevalues[state]=0
continue
actionvalue = -100000
for action in mdp.getPossibleActions(state):
S_next = self.getQValue(state,action)
if actionvalue<S_next:
actionvalue = S_next
statevalues[state] = actionvalue
self.values = statevalues
"*** 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 ***"
allStates = mdp.getStates()
vPrimes = util.Counter() # A Counter is a dict with default 0
iteration = 0
while iteration < iterations:
for s in allStates:
if mdp.isTerminal(s):
vPrimes[s] = mdp.getReward(s, None, s);
else:
sreward = mdp.getReward(s, None, s)
vPrimes[s] = sreward + discount * self.utilOfBestAction(mdp, s )
for s in allStates:
self.values[s] = vPrimes[s]
iteration +=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 ***"
#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)
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
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
for i in xrange(iterations):
tmp = util.Counter()
for state in mdp.getStates():
tmp[state] = self.values[state]
for state in mdp.getStates():
if not mdp.isTerminal(state):
q_value = -10000000
for action in mdp.getPossibleActions(state):
q_value = max(q_value, self.computeQValueFromValues(state, action))
tmp[state] = q_value
self.values = tmp
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
for loop in range(self.iterations):
value_dict = util.Counter()
mdp_list = mdp.getStates()
for state in mdp_list:
#check if the agents isTerminal
if not mdp.isTerminal(state):
#make counter_dict and tuple for possible actions
counter_dict = util.Counter()
actions_tuple = mdp.getPossibleActions(state)
#iterate and get qValue
for action in actions_tuple:
counter_dict[action] = self.computeQValueFromValues(
state, action)
#get max of the counter_dict values
value_dict[state] = max(counter_dict.values())
#make a copy and update values
self.values = value_dict.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
"*** 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
newValues = util.Counter()
# Write value iteration code here
"*** YOUR CODE HERE ***"
allStates = mdp.getStates()
for i in range(iterations):
for s in allStates:
if mdp.isTerminal(s): continue
mx = float("-inf")
for a in mdp.getPossibleActions(s):
score = 0
for (sp, tp) in mdp.getTransitionStatesAndProbs(s, a):
score += tp * (mdp.getReward(s, a, sp) +
self.discount * self.values[sp])
if score > mx:
mx = score
newValues[s] = mx
self.values = newValues.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
self.tempvalues = self.values.copy()
# Write value iteration code here
#print "start location ", mdp.getStates()
"*** YOUR CODE HERE ***"
for i in xrange(0, self.iterations):
for allstate in mdp.getStates():
if not mdp.isTerminal(allstate):
value = []
for possibleaction in mdp.getPossibleActions(allstate):
statevalue = 0
statevalue = statevalue + self.computeQValueFromValues(
allstate, possibleaction)
value.append(statevalue)
self.tempvalues[allstate] = max(value)
self.values = self.tempvalues.copy()
def computeActionFromValues(self, state):
"""
The policy is the best action in the given state
according to the values currently stored in self.values.
You may break ties any way you see fit. Note that if
there are no legal actions, which is the case at the
terminal state, you should return None.
"""
"*** YOUR CODE HERE ***"
mdp = self.mdp
possibleActions = mdp.getPossibleActions(state)
maxActionValue = float('-inf')
maxAction = None
if ((possibleActions == None) or (mdp.isTerminal(state))):
return None
for action in possibleActions:
actionSum = self.getQValue(state, action)
#Find the maximum action
if maxActionValue < actionSum:
maxAction = action
maxActionValue = actionSum
return maxAction
def runValueIteration(self):
# Write value iteration code here
# print(self.iterations)
for i in range(0, self.iterations):
newValues = {}
for state in self.values:
newValues[state] = [state]
# print("iteartion ",i)
mdp = self.mdp
statesInMDP = mdp.getStates()
for currentState in statesInMDP:
if (not mdp.isTerminal(currentState)):
actions = mdp.getPossibleActions(currentState)
maxActionReward = -math.inf
for currentAction in actions:
currentActionReward = self.computeQValueFromValues(
currentState, currentAction)
if (currentActionReward > maxActionReward):
maxActionReward = currentActionReward
newValues[currentState] = maxActionReward
else:
newValues[currentState] = self.getValue(currentState)
self.values = newValues
# mdp.isTerminal(state)
"*** 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 i in range(0, iterations):
new_values = copy.copy(self.values)
for s in mdp.getStates():
if mdp.isTerminal(s):
new_values[s] = 0.0
else:
values = []
actions = mdp.getPossibleActions(s)
for anaction in actions:
values.append(self.getQValue(s, anaction))
new_values[s] = max(values)
self.values = new_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)
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 computeActionFromValues(self, state):
"""
The policy is the best action in the given state
according to the values currently stored in self.values.
You may break ties any way you see fit. Note that if
there are no legal actions, which is the case at the
terminal state, you should return None.
"""
"*** YOUR CODE HERE ***"
import random
mdp = self.mdp
possibleActions = mdp.getPossibleActions(state)
valuesForAction = util.Counter()
if (mdp.isTerminal(state)):
return None
for action in possibleActions:
tp = mdp.getTransitionStatesAndProbs(state, action)
sum = 0
for i in range(len(tp)):
nextState, prob = tp[i]
sum = sum + prob * (mdp.getReward(state, action, nextState) +
self.discount * self.values[nextState])
valuesForAction[action] = sum
if (valuesForAction.totalCount() == 0):
return possibleActions[0]
return valuesForAction.argMax()
def computeActionFromValues(self, state):
"""
The policy is the best action in the given state
according to the values currently stored in self.values.
You may break ties any way you see fit. Note that if
there are no legal actions, which is the case at the
terminal state, you should return None.
"""
"*** YOUR CODE HERE ***"
mdp = self.mdp
possibleActions = mdp.getPossibleActions(state)
maxActionValue = float('-inf')
maxAction = None
if ((possibleActions==None) or (mdp.isTerminal(state))):
return None
for action in possibleActions:
actionSum = self.getQValue(state, action)
#Find the maximum action
if maxActionValue < actionSum:
maxAction = action
maxActionValue = actionSum
return maxAction
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
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
for i in range(self.iterations):
valCopy = self.values.copy()
for state in self.mdp.getStates():
maxVal = float("-inf")
if not mdp.isTerminal(state):
for action in mdp.getPossibleActions(state):
maxVal = max(
maxVal,
self.computeQValueFromValues(state, action))
valCopy[state] = maxVal
self.values = valCopy
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 ***"
temp = util.Counter()
for i in xrange(iterations):
temp = self.values.copy()
for j in mdp.getStates():
vlist = []
actions = mdp.getPossibleActions(j)
if not mdp.isTerminal(j):
for k in actions:
tran = mdp.getTransitionStatesAndProbs(j, k)
val = 0
for m in tran:
val += m[1] * (mdp.getReward(j, k, m[0]) + self.discount * temp[m[0]])
vlist.append(val)
self.values[j] = max(vlist)
def computeActionFromValues(self, state):
"""
The policy is the best action in the given state
according to the values currently stored in self.values.
You may break ties any way you see fit. Note that if
there are no legal actions, which is the case at the
terminal state, you should return None.
"""
"*** YOUR CODE HERE ***"
mdp = self.mdp
max_val = -100
direction = "exit"
if mdp.isTerminal(state) == True:
return "exit"
else:
for action in mdp.getPossibleActions(state):
Q_value = self.computeQValueFromValues(state, action)
if max_val < Q_value:
max_val = Q_value
direction = action
'''max_val = Q_value if max_val < Q_value else max_val
direction = action if max_val < Q_value else None'''
return direction
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):
oldv = self.values.copy()
for s in mdp.getStates():
if (mdp.isTerminal(s)):
self.values[s] = 0
continue
maxu = None
for a in mdp.getPossibleActions(s):
eu = 0
for (sp, p) in mdp.getTransitionStatesAndProbs(s, a):
r = mdp.getReward(s, a, sp)
r += self.discount * oldv[sp]
eu += p * r
if (maxu is None or eu > maxu): maxu = eu
self.values[s] = maxu
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 = self.mdp.getStates()
for i in range(self.iterations):
tempValue = util.Counter()
for s in states:
if mdp.isTerminal(s):
val = 0
tempValue[s] = val
else:
val = float("-inf")
for a in self.mdp.getPossibleActions(s):
temp_value = self.computeQValueFromValues(s, a)
val = max(temp_value, val)
tempValue[s] = val
self.values = tempValue
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
next_values = util.Counter()
for i in range(0, iterations):
for state in mdp.getStates():
if not mdp.isTerminal(state):
arr = []
for action in self.mdp.getPossibleActions(state):
arr.append(self.getQValue(state, action))
next_values[state] = max(arr)
self.values = next_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)
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 index in range(self.iterations):
nextValues = util.Counter()
for state in mdp.getStates():
if mdp.isTerminal(state):
# terminal states have 0 value;
# their actual 'value' is the transition reward
continue
qValues = []
for action in mdp.getPossibleActions(state):
qValues.append(self.computeQValueFromValues(state, action))
nextValues[state] = max(qValues)
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.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
# 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
"*** YOUR CODE HERE ***"
states = mdp.getStates()
#Iterate through user-defined number of iterations
for num in range(iterations):
temp = util.Counter()
#Compute Ut+1 for all states
for state in states:
if mdp.isTerminal(state):
self.values[state] = 0
continue
actions = mdp.getPossibleActions(state)
maxVal = MIN
#iterate through trans of each action of the state and sum up values
for action in actions:
transitions = mdp.getTransitionStatesAndProbs(state, action)
totalSum = 0
for transition in transitions:
#transition[0] = nextState, transition[1] = probability
reward = mdp.getReward(state, action, transition[0])
#value of the nextState
UtValue = self.values[transition[0]]
#using formula of value iteration from wikipedia
totalSum += transition[1]*(reward + discount * UtValue)
maxVal = max(maxVal, totalSum)
#for some reason, self.values[state] = maxVal doesn't work.
temp[state] = maxVal
for state in states:
self.values[state] = temp[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 ***"
oldValues = util.Counter()
self.policy = util.Counter()
states=self.mdp.getStates()
for k in range (iterations):
for state in states:
actionValues = util.Counter()
actions=self.mdp.getPossibleActions(state)
for action in actions:
aux=self.mdp.getTransitionStatesAndProbs(state, action)
for things in aux:
#print "nextState:",things[0]
#print "prob:",things[1]
actionValues[action]+=things[1]*(mdp.getReward(state,action,things[0])+self.discount*oldValues[things[0]])
tmp = actionValues.argMax()
self.values[state] = actionValues[tmp]
if mdp.isTerminal(state):
self.policy[state] = None
else:
self.policy[state] = tmp
oldValues=self.values.copy()
def computeValue(self,mdp,state,discount):
actions = mdp.getPossibleActions(state)
valueList = []
if (mdp.isTerminal(state)):
return
for action in actions:
transitions = mdp.getTransitionStatesAndProbs(state,action)
value = 0
for transition in transitions:
subValue = float(transition[1]) * (float(mdp.getReward(state,action,transition[0]))+(float(discount)*(float(self.getValue(transition[0])))))
value += subValue
valueList.append(value)
self.tmpValues[state] = max(valueList)
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)
mdp.isTerminal(state)
"""
self.mdp = mdp
self.discount = discount
self.iterations = iterations
self.values = util.Counter() # A Counter is a dict with default 0
# NOTE: The naming conventions are a bit off. This is to emphasize the
# connection between code and the formulas we were provided with.
# Enjoy!
import sys
values_temp = util.Counter()
for i in range(iterations):
for state in mdp.getStates():
Q_max = -sys.maxint
# A terminal state has no actions, so we must be careful to
# reset the value to zero here.
if mdp.isTerminal(state):
Q_max = 0.0
# This is a trivial loop to find the 'best' possible action in
# the current state, according to computed Q values. This is
# essentially the Pythonic way of saying the following:
# V_k+1(s) <- max Q(s, a)
for action in mdp.getPossibleActions(state):
Q = self.getQValue(state, action)
if Q > Q_max:
Q_max = Q
values_temp[state] = Q_max
# Store the new values.
self.values = values_temp
values_temp = util.Counter()
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 "Final state data"
# print self.mdp.getPossibleActions((3,1))
# print self.mdp.getTransitionStatesAndProbs((3,1),"exit")
# print "Final state data"
for i in range(self.iterations):
newValues = self.values.copy()
for state in self.mdp.getStates():
# print "state is "+ str(state)
if (mdp.isTerminal(state)):
continue
possibleActions = self.mdp.getPossibleActions(state)
# print possibleActions
maxVal=-999
for action in possibleActions:
# print action
val = self.computeQValueFromValues(state,action)
# print val
if val>maxVal:
maxVal = val
newValues[state] = maxVal
self.values = newValues.copy()
def __init__(self, mdp, discount = 0.9, iterations = 1000):
"""
Your cyclic value iteration agent should take an mdp on
construction, run the indicated number of iterations,
and then act according to the resulting policy. Each iteration
updates the value of only one state, which cycles through
the states list. If the chosen state is terminal, nothing
happens in that iteration.
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 = collections.defaultdict(float)
states = self.mdp.getStates()
for state in states:
self.values[state] = 0
"*** YOUR CODE HERE ***"
for i in range(iterations):
start = time.time()
state = states[i%len(states)]
if not mdp.isTerminal(state):
reward = mdp.getReward(state)
actions = mdp.getPossibleActions(state)
EUActions = []
for action in actions:
EUAction = 0
transitions = mdp.getTransitionStatesAndProbs(state, action)
for transition in transitions:
transitionState = transition[0]
transitionUtility = self.values[transitionState]
transitionProbability = transition[1]
EUAction += transitionUtility*transitionProbability
EUActions.append(EUAction)
maxEU = max(EUActions)
updatedUtility = reward + discount*maxEU
self.values[state] = updatedUtility
print('Iteration: ' + str(i))
print(sum(abs(value - 100) for state, value in self.values.items() if not self.mdp.isTerminal(state)))
print('Time elapsed: ' + str(time.time()-start))
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.action_array = {}
self.last_values = util.Counter()
# Write value iteration code here
"*** YOUR CODE HERE ***"
for state in mdp.getStates():
self.values[state] = 0.0
while self.iterations > 0:
for state in mdp.getStates():
qact = {}
if mdp.isTerminal(state):
continue
else:
currentActions = mdp.getPossibleActions(state)
for action in currentActions:
qact[action] = self.computeQValueFromValues(state, action)
#print action, qact.values()
#for v in qact.values():
# tmp.append(v)
self.values[state] = max(qact.values())
for k, v in qact.items():
#print state
#print k, v
if v == self.values[state]:
self.action_array[state] = k
self.last_values = self.values.copy()
#print self.last_values
self.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
# Write value iteration code here
"*** YOUR CODE HERE ***"
Unew = util.Counter()
delta = 0
for iterate in range(0,self.iterations):
Unew = util.Counter()
for states in mdp.getStates():
if mdp.isTerminal(states):
Unew[states] = 0
continue;
actions = mdp.getPossibleActions(states)
if not actions:
Unew[states] = 0
maxQvalue = float("-inf")
qValue = 0
for a in actions:
qValue = self.getQValue(states,a);
if qValue > maxQvalue:
maxQvalue = qValue
Unew[states] = qValue
self.values = Unew
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 ***"
import copy
for iteration in range(iterations):
newvalues = util.Counter()
for state in mdp.getStates():
if mdp.isTerminal(state):
newvalues[state] = 0
continue
possibleActions = mdp.getPossibleActions(state)
maxAction = None
maxActionValue = float('-inf')
if (possibleActions==None):
newvalues[state] = 0
for action in possibleActions:
actionsum = self.getQValue(state, action)
#Find the maximum action
if maxActionValue < actionsum:
maxAction = action
maxActionValue = actionsum
#maxActionValue is now V:k+1 after iteration
newvalues[state] = maxActionValue
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 ***"
# State = Coordinate
for iteration in range(iterations):
newValues = util.Counter()
oldValues = self.values.copy()
allStates = mdp.getStates()
for state in allStates:
if not mdp.isTerminal(state):
rewardsForActions = util.Counter()
possibleActions = mdp.getPossibleActions(state)
for possibleAction in possibleActions:
transitionStatesAndProbs = mdp.getTransitionStatesAndProbs(state, possibleAction)
for transitionStateAndProb in transitionStatesAndProbs:
newState = transitionStateAndProb[0]
prob = transitionStateAndProb[1]
rewardForTransition = mdp.getReward(state, possibleAction, newState)
rewardsForActions[possibleAction] = rewardsForActions[possibleAction] + prob * (
rewardForTransition + (discount * oldValues[newState])
)
maxAction = rewardsForActions.argMax()
newValues[state] = rewardsForActions[maxAction]
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
counter = 0
for state in self.mdp.getStates():
self.values[state] = 0
while counter < iterations:
newValues = util.Counter()
for state in self.mdp.getStates():
if not mdp.isTerminal(state):
actions = self.mdp.getPossibleActions(state)
maxValue = float("-inf")
for action in actions:
value = 0
stateProbs = self.mdp.getTransitionStatesAndProbs(state, action)
for pair in stateProbs:
discounted = 0
if pair[0] == 'TERMINAL_STATE':
discounted = 0 #redundant???
else:
discounted = discount * self.getValue(pair[0])
value += (self.mdp.getReward(state, action, pair[0]) + discounted) * pair[1]
if value > maxValue:
maxValue = value
newValues[state] = maxValue
for state in newValues:
self.values[state] = newValues[state]
counter += 1
print self.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 "-----------------------------------------------------"
"*** MY CODE BEGINS ***"
k = 0
while k < iterations:
val = self.values.copy() #before each iteration, copy one.
for s in mdp.getStates():
if mdp.isTerminal(s) == False:
max = -999999
for action in mdp.getPossibleActions(s):
v = 0
for pos_pro in mdp.getTransitionStatesAndProbs(s,action):
v = v + pos_pro[1]*(mdp.getReward(s,action,pos_pro[0])+discount*self.values[pos_pro[0]])
if v > max:
max = v
val[s] = max
else:
for action in mdp.getPossibleActions(s):
v = 0
for pos_pro in mdp.getTransitionStatesAndProbs(s,action):
v = v + pos_pro[1]*(mdp.getReward(s,action,pos_pro[0])+discount*self.values[pos_pro[0]])
val[s] = v
k = k+1
for s in mdp.getStates():
self.values[s] = val[s]