def generate_mdp(mdp_type):
if mdp_type == 'tunnel':
return Mdp.ChainsTunnelMDP(mdpcfg.TunnelMDPConfig())
if mdp_type == 'star':
return Mdp.StarMDP(mdpcfg.StarMDPConfig())
if mdp_type == 'clique':
return Mdp.CliquesMDP(mdpcfg.CliqueMDPConfig())
if mdp_type == 'cliff':
return Mdp.CliffWalker(mdpcfg.CliffMDPConfig())
if mdp_type == 'directed':
return Mdp.DirectedTreeMDP(
mdpcfg.DirectedTreeMDPConfig())
raise NotImplementedError()
CONF_FILE = TIRAMOLA_DIR + "complex_20k/mdp_small.json"
#############################################################
PRINT = len(sys.argv) > 1 and sys.argv[1] == "-p"
if PRINT:
num_tests = 1
conf = ModelConf(CONF_FILE)
assert conf.get_model_type() == MDP, "Wrong model type in MDP example"
total_reward_results = []
for i in range(num_tests):
scenario = ComplexScenario(training_steps, load_period, 10, MIN_VMS, MAX_VMS)
model = MDPModel(conf.get_model_conf())
model.set_state(scenario.get_current_measurements())
total_reward = 0
for time in range(training_steps + eval_steps):
if random.uniform(0, 1) < epsilon and time < training_steps:
action = random.choice(model.get_legal_actions())
else:
action = model.suggest_action()
reward = scenario.execute_action(action)
meas = scenario.get_current_measurements()
model.update(action, meas, reward)
if time % 500 == 1:
reward = served_load - 3 * vms
return reward
def average(l):
return sum(l) / len(l)
conf = ModelConf(CONF_FILE)
assert conf.get_model_type() == MDP, "Wrong model type in MDP example"
total_reward_results = []
for i in range(num_tests):
model = MDPModel(conf.get_model_conf())
m = get_next_measurements({NUMBER_OF_VMS: 10}, (NO_OP, 0), 0)
model.set_state(m)
total_reward = 0
for time in range(training_steps + eval_steps):
if random.uniform(0, 1) < epsilon and time < training_steps:
action = random.choice(model.get_legal_actions())
else:
action = model.suggest_action()
m = get_next_measurements(m, action, time)
reward = get_reward(m, action)
model.update(action, m, reward)
TIRAMOLA_DIR = "/home/kostis/git/tiramola/"
import sys
sys.path.append(TIRAMOLA_DIR)
from Configuration import ModelConf
from Constants import *
from MDPModel import MDPModel
CONFIGURATION_FILE = TIRAMOLA_DIR + "examples/mdp_basic/mdp_basic.json"
conf = ModelConf(CONFIGURATION_FILE)
assert conf.get_model_type() == MDP, "Wrong model type in MDP example"
model_conf = conf.get_model_conf()
model = MDPModel(model_conf)
m1 = {TOTAL_LOAD: 0.2534, NUMBER_OF_VMS: 7, TOTAL_LATENCY: 0.156}
m2 = {TOTAL_LOAD: 1.5523, NUMBER_OF_VMS: 5, TOTAL_LATENCY: 0.524}
m3 = {TOTAL_LOAD: 1.6605, NUMBER_OF_VMS: 6, TOTAL_LATENCY: 0.100}
m4 = {TOTAL_LOAD: 0.1005, NUMBER_OF_VMS: 6, TOTAL_LATENCY: 0.250}
model.set_prioritized_sweeping(0.01, 100)
model.set_state(m1)
action = (REMOVE_VMS, 1)
model.update(action, m2, 2.5)
model.update(action, m3, 1.4)
model.update(action, m4, 4.2)
model.print_model(detailed=True)
#model.value_iteration(0.01)
def get_reward(measurements, time, action):
load = get_load(time)
vms = measurements[NUMBER_OF_VMS]
reward = min(10 * vms, load) - 3 * vms
if action[0] == ADD_VMS:
reward -= 2 * action[1]
elif action[0] == REMOVE_VMS:
reward -= action[1]
return reward
conf = ModelConf(TIRAMOLA_DIR + "examples/mdp_2d_ps/mdp_2d_ps.json")
assert conf.get_model_type() == MDP, "Wrong model type in MDP example"
model = MDPModel(conf.get_model_conf())
max_steps = 5000
vi_step = 2500
epsilon = 0.3
time = 0
measurements = {NUMBER_OF_VMS: 1, TOTAL_LOAD: get_load(0)}
model.set_state(measurements)
while time < max_steps:
time += 1
if random.uniform(0, 1) < epsilon and time < vi_step:
action = random.choice(model.get_legal_actions())
else: