def actionablestatesfrom(state):
global maxdisseen
t = tfs(state)
nextdisvalues = np.round([
modelt4.cocoef * disstates[dfs(state)] + modelt4.phi(t, customer, a)
for a in actions if a >= modelt4.wholeprice(t)
], -2)
maxdisseen = max(maxdisseen, max(nextdisvalues))
nextdisvalues = [min(value, disstates[-1]) for value in nextdisvalues]
idx = getdisstate(nextdisvalues)
return list(set([(t) + i * (ntimeslots + 1) for i in idx]))
def takeaction(state,n,greedy = True):
# if greedy and np.random.random() <= min(epsilon, 10/iterations):
if greedy and np.random.random() <= min(epsilon, 1000/iterations):
price = 0
while price < modelt4.wholeprice(tfs(state)):
randomaction = np.random.randint(nactions)
price = actions[randomaction]
#randomprice = actions[randomaction]
return randomaction
else:
bestaction = np.argmax(qmatrix[state,:])
#bestprice = actions[bestaction]
return bestaction
# initialization
timeslot = 1
iterations = 0
qmatrix = np.zeros([nstates, nactions]) # one extra row
qprev = -100 * np.ones([nstates, nactions]) # one extra row
delta = 0.1
convergence = []
qconvergence = []
rewardgraph = []
initstate = getstate(1, initdis)
#qmatrix = np.full([ntimeslots+1,nactions], -np.inf) # one extra row
for i in range(nstates):
for j in range(nactions):
if tfs(i) <= ntimeslots:
if actions[j] < modelt4.wholeprice(tfs(i)):
qprev[i, j] = -1000
qmatrix[i, j] = -100
def currentpolicy():
bpolicy = list()
policystate = initstate
action = np.argmax(qmatrix[policystate, :])
price = actions[action]
bpolicy.append(price)
#policystate = np.argmax(qmatrix[:,:], axis=1)
for t in range(2, 25):
policystate = nextstate(policystate, price)
action = np.argmax(qmatrix[policystate, :])
price = actions[action]