def __init__(self,
host="localhost",
port=6379,
db=0,
freq_sim=1000,
freq_control=100):
self.links = ()
self.rbdl = Rbdl()
self.simulator = Simulator(self.rbdl,
freq_sim=freq_sim,
freq_control=freq_control,
asynchronous=False)
self.redis_db = redis.Redis(host=host,
port=port,
db=db,
decode_responses=True)
self.publish_cv = threading.Condition()
self.publish_queue = None
self.publisher = threading.Thread(target=self.publisher_thread)
self.publisher.daemon = True
self.publisher.start()
self.set_gravity()
self.set_ee_offset()
self.set_joint_limits()
def simulate(self):
'''
This method is used to solve the resulting initial value problem
after the computation of a solution for the input trajectories.
'''
logging.debug("Solving Initial Value Problem")
# calulate simulation time
T = self.dyn_sys.b - self.dyn_sys.a
# get list of start values
start = []
if self.constraints is not None:
sys = self._dyn_sys_orig
else:
sys = self.dyn_sys
x_vars = sys.states
start_dict = dict([(k, v[0]) for k, v in sys.boundary_values.items()
if k in x_vars])
ff = sys.f_num
for x in x_vars:
start.append(start_dict[x])
# create simulation object
S = Simulator(ff, T, start, self.eqs.trajectories.u)
logging.debug("start: %s" % str(start))
# start forward simulation
self.sim_data = S.simulate()
def btnStartPauseClicked(self):
"""
Start/Pause button is clicked
"""
if self.data['Geno'] is None:
# data not loaded yet
print('Data not loaded yet')
self.changeStatus('Please load the data first.')
else:
# start simulation
if self.simulator is None:
# if it is the first time to start the simulation
self.simulator = Simulator(self.data, self.gridLayout, self.canvas, self.listStatus, self.listSimStatus)
self.simulator.start()
self.btnStartPause.setText('Pause')
else:
if self.simulator.isSimStopped():
self.simulator.restart()
else:
if self.simulator.isSimPaused():
self.simulator.resume()
self.btnStartPause.setText('Pause')
else:
self.simulator.pause()
self.btnStartPause.setText('Resume')
def run_n_simulations(designpoint,
dp_name,
iterations,
outputs,
all_samples=False):
""" Evaluate the given design point by calculating the MTTF for a given amount of sample_budget.
This function is used for parallelising the Monte Carlo Simulation evaluation.
:param designpoint: DesignPoint object - this DesignPoint is evauluated.
:param dp_name: any - unique inidcator of this designpoint (e.g. integer)
:param iterations: number of sample_budget to run the MCS.
:param outputs: dictionary to write the MTTF output.
:return: None
"""
TTFs = []
consumptions = []
sim = Simulator(designpoint)
for i in range(iterations):
ttf, consum, size = sim.run()
TTFs.append(ttf)
consumptions.append(consum)
if not all_samples:
outputs[dp_name] = sum(TTFs) / len(TTFs), sum(consumptions) / len(
consumptions), size
else:
outputs[dp_name] = list(
zip(TTFs, consumptions, [size for _ in range(len(TTFs))]))
def simulate(self):
'''
This method is used to solve the resulting initial value problem
after the computation of a solution for the input trajectories.
'''
logging.debug("Solving Initial Value Problem")
# calulate simulation time
T = self.dyn_sys.b - self.dyn_sys.a
# get list of start values
start = []
if self.constraints is not None:
sys = self._dyn_sys_orig
else:
sys = self.dyn_sys
x_vars = sys.states
start_dict = dict([(k, v[0]) for k, v in sys.boundary_values.items() if k in x_vars])
ff = sys.f_num
for x in x_vars:
start.append(start_dict[x])
# create simulation object
S = Simulator(ff, T, start, self.eqs.trajectories.u)
logging.debug("start: %s"%str(start))
# start forward simulation
self.sim_data = S.simulate()
def compareChurnInjection():
s = Simulator()
random.seed(125)
s.setupSimulation(strategy="churn",
workMeasurement="one",
numNodes=1000,
numTasks=100000,
churnRate=0.01)
loads1, medians1, means1, maxs1, devs1 = s.simulateLoad()
random.seed(125)
s = Simulator()
s.setupSimulation(strategy='randomInjection',
workMeasurement="one",
numNodes=1000,
numTasks=100000,
churnRate=0)
loads2 = s.simulateLoad()[0]
for i in range(0, len(loads1), 5):
x1 = loads1[i]
x2 = loads2[i]
colors = ["r", "b"]
labels = ["Churn", "Random Injection"]
plt.hist([x1, x2], 25, normed=1, color=colors, label=labels)
plt.legend(loc=0)
plt.title('Churn vs Random Injection at Tick ' + str(i))
plt.xlabel('Tasks Per Node')
plt.ylabel('Fraction of the Network')
#plt.ylim(0, 0.05)
plt.show()
def compareInjectionStable():
s = Simulator()
random.seed(125)
s.setupSimulation(strategy="churn",
homogeneity="equal",
workMeasurement="one",
numNodes=1000,
numTasks=100000,
churnRate=0)
loads1, medians1, means1, maxs1, devs1 = s.simulateLoad()
random.seed(125)
s = Simulator()
s.setupSimulation(strategy="randomInjection",
homogeneity="perStrength",
workMeasurement="one",
numNodes=1000,
numTasks=100000,
churnRate=0)
loads2 = s.simulateLoad()[0]
for i in range(0, len(loads1), 5):
x1 = loads1[i]
x2 = loads2[i]
colors = ["k", "w"]
labels = ["No Strategy", "Random Injection"]
plt.hist([x1, x2], 25, normed=1, color=colors, label=labels)
plt.legend(loc=0)
plt.title('Random Injection in a Heterogeneous Network at Tick ' +
str(i))
plt.xlabel('Tasks Per Node')
plt.ylabel('Fraction of the Network')
#plt.ylim(0, 0.05)
plt.show()
def compareInviteNeighbor():
s = Simulator()
random.seed(125)
s.setupSimulation(strategy="neighbors",
workMeasurement="one",
numNodes=1000,
numTasks=100000,
churnRate=0)
loads1, medians1, means1, maxs1, devs1 = s.simulateLoad()
random.seed(125)
s = Simulator()
s.setupSimulation(strategy="invite",
workMeasurement="one",
numNodes=1000,
numTasks=100000,
churnRate=0)
loads2 = s.simulateLoad()[0]
for i in range(0, len(loads1), 5):
x1 = loads1[i]
x2 = loads2[i]
colors = ["k", "w"]
labels = ["Neighbors", "Invitation"]
plt.hist([x1, x2], 25, normed=1, color=colors, label=labels)
plt.legend(loc=0)
plt.title('Invitation vs Smart Neighbors at Tick ' + str(i))
plt.xlabel('Tasks Per Node')
plt.ylabel('Fraction of the Network')
#plt.ylim(0, 0.05)
plt.show()
def validate_MDP_policy(root_path, flag_with_obs=True, flag_plan=True):
s_g = np.array([4.0, 3.0, 0.0])
modality = "haptic"
obs_list = [(1.5, 2.00, 1.0, 1.25),
(2.0, 1.0, 1.25, 0.5)]
if flag_with_obs:
file_name = root_path + "/mdp_planenr_obs_" + modality + ".pkl"
else:
file_name = root_path + "/mdp_planner_" + modality + ".pkl"
if flag_plan:
human_model = MovementModel()
human_model.load_model(root_path)
human_model.set_default_param()
planner = MDPFixedTimePolicy(tmodel=human_model)
if flag_with_obs:
planner.gen_env(obs_list)
planner.compute_policy(s_g, modality, max_iter=30)
else:
with open(file_name) as f:
planner = pickle.load(f)
fig, axes = plt.subplots()
planner.visualize_policy(axes)
plt.show()
if flag_plan:
with open(file_name, "w") as f:
pickle.dump(planner, f)
sim = Simulator(planner)
n_trials = 30
traj_list = []
start_time = time.time()
for i in range(n_trials):
traj_list.append(sim.run_trial((0.5, 0.5, 0.0), s_g, modality, 30.0, tol=0.5))
print "--- %s seconds ---" % (time.time() - start_time)
fig, axes = plt.subplots()
for i in range(n_trials):
t, traj = traj_list[i]
axes.plot(traj[:, 0], traj[:, 1])
axes.axis("equal")
axes.scatter(s_g[0], s_g[1])
if flag_with_obs:
for x, y, w, h in obs_list:
rect = Rectangle((x, y), w, h)
axes.add_patch(rect)
plt.show()
def validate_MDP_policy(root_path, flag_with_obs=True, flag_plan=True):
s_g = np.array([4.0, 3.0, 0.0])
modality = "haptic"
obs_list = [(1.5, 2.00, 1.0, 1.25), (2.0, 1.0, 1.25, 0.5)]
if flag_with_obs:
file_name = root_path + "/mdp_planenr_obs_" + modality + ".pkl"
else:
file_name = root_path + "/mdp_planner_" + modality + ".pkl"
if flag_plan:
human_model = MovementModel()
human_model.load_model(root_path)
human_model.set_default_param()
planner = MDPFixedTimePolicy(tmodel=human_model)
if flag_with_obs:
planner.gen_env(obs_list)
planner.compute_policy(s_g, modality, max_iter=30)
else:
with open(file_name) as f:
planner = pickle.load(f)
fig, axes = plt.subplots()
planner.visualize_policy(axes)
plt.show()
if flag_plan:
with open(file_name, "w") as f:
pickle.dump(planner, f)
sim = Simulator(planner)
n_trials = 30
traj_list = []
start_time = time.time()
for i in range(n_trials):
traj_list.append(
sim.run_trial((0.5, 0.5, 0.0), s_g, modality, 30.0, tol=0.5))
print "--- %s seconds ---" % (time.time() - start_time)
fig, axes = plt.subplots()
for i in range(n_trials):
t, traj = traj_list[i]
axes.plot(traj[:, 0], traj[:, 1])
axes.axis("equal")
axes.scatter(s_g[0], s_g[1])
if flag_with_obs:
for x, y, w, h in obs_list:
rect = Rectangle((x, y), w, h)
axes.add_patch(rect)
plt.show()
def __init__(self):
self.dt = 0.01
self.max_force = 125
self.step_n = int(450) # number of steps
self.simulator = Simulator()
self.Ctheta = -1
self.Ktheta = 0.05
self.Cx = 0.5
self.Kx = 10
def testChurnSteps():
s = Simulator()
s.setupSimulation(strategy='churn',
workMeasurement="one",
numNodes=1000,
numTask=100000,
churnRate=0.001)
loads, medians, means, maxs, devs = s.simulateLoad()
print(medians)
print(devs)
def runSim(trials):
sim = Simulator(requests)
outFile = open("data/results/{0}_{1}_{2}_delays.results".format(week, day, trials), 'w')
# Report the results to the open file.
def reportResults(report):
report.printTSV(outFile)
if trials > 1:
runSim(trials - 1)
# Run the simulation.
tieBreakingPolicy = crisisThresholdsAndFlip(10 * 60, 40 * 60, 1.5)
sim.run(reportResults, tieBreakingPolicy, regular, senior)
def testInjectionSteps():
s = Simulator()
s.setupSimulation(strategy = 'randomInjections',workMeasurement="one", numNodes=1000, numTasks=100000)
loads, medians, means, maxs, devs = s.simulateLoad()
for i in range(0,len(loads), 5):
x = loads[i]
plt.hist(x, 100, normed =1 )
plt.xlabel('Tasks Per Node')
plt.ylabel('Probability')
plt.axvline(medians[i], color='r', linestyle='--')
plt.axvline(means[i], color='k', linestyle='--')
#plt.ylim(0, 0.05)
plt.show()
def save_timing_performance(timing, filename):
full_timings = [3] * 8
full_timings[0] = timing[0]
full_timings[4] = timing[1]
traffic_light = PhaseModifier("node1")
controller = StaticTrafficLightController(controller=traffic_light,
sequence=list(range(8)),
timings=full_timings)
sim = Simulator()
sim.add_simulation_component(SimulationOutputParser)
sim.add_tickable(controller)
sim.run(sumocfg1, gui=False)
sim.save_results(filename)
def plotLoads():
s = Simulator()
seed = 500
loads = []
for _ in range(20):
random.seed(seed)
s.setupSimulation(numNodes=1000,numTasks=1000000)
loads = loads + [len(x.tasks) for x in s.nodes.values()]
seed += 1
n, bins, patches = plt.hist(loads, 25, normed =1 )
plt.xlabel('Tasks Per Node')
plt.ylabel('Probability')
plt.axvline(statistics.median_low(loads), color='r', linestyle='--')
plt.show()
def plotLoads():
s = Simulator()
seed = 500
loads = []
for _ in range(20):
random.seed(seed)
s.setupSimulation(numNodes=1000, numTasks=1000000)
loads = loads + [len(x.tasks) for x in s.nodes.values()]
seed += 1
n, bins, patches = plt.hist(loads, 25, normed=1)
plt.xlabel('Tasks Per Node')
plt.ylabel('Probability')
plt.axvline(statistics.median_low(loads), color='r', linestyle='--')
plt.show()
def monte_carlo_iterative(designpoints, sample_budget, all_samples=False):
""" Iterative implementation of the MCS to rank the given design points.
:param designpoints: [DesignPoint object] - List of designpoint objects (the candidates).
:param sample_budget: number of MC sample_budget to run
:return: [float] - List of MTTF corresponding indexwise to the design points.
"""
TTFs = {i: [] for i in range(len(designpoints))}
consumptions = {i: [] for i in range(len(designpoints))}
sims = [Simulator(d) for d in designpoints]
i_per_dp = sample_budget // len(designpoints)
for i in range(len(designpoints)):
for _ in range(i_per_dp):
ttf, consum, size = sims[i].run()
TTFs[i].append(ttf)
consumptions[i].append(consum)
for i in range(len(TTFs)):
TTFs[i] = sum(TTFs[i]) / len(TTFs[i])
consumptions[i] = sum(consumptions[i]) / len(consumptions[i])
output = {i: [] for i in range(len(designpoints))}
for i in TTFs:
output[i] = (TTFs[i], consumptions[i], designpoints[i].evaluate_size())
return output
def testInjectionSteps():
s = Simulator()
s.setupSimulation(strategy='randomInjections',
workMeasurement="one",
numNodes=1000,
numTasks=100000)
loads, medians, means, maxs, devs = s.simulateLoad()
for i in range(0, len(loads), 5):
x = loads[i]
plt.hist(x, 100, normed=1)
plt.xlabel('Tasks Per Node')
plt.ylabel('Probability')
plt.axvline(medians[i], color='r', linestyle='--')
plt.axvline(means[i], color='k', linestyle='--')
#plt.ylim(0, 0.05)
plt.show()
def siumlate_with_input(tp, inputseq, n_parts ):
"""
:param tp: TransitionProblem
:param inputseq: Sequence of input values (will be spline-interpolated)
:param n_parts: number of spline parts for the input
:return:
"""
tt = np.linspace(tp.a, tp.b, len(inputseq))
# currently only for single input systems
su1 = new_spline(tp.b, n_parts, (tt, inputseq), 'u1')
sim = Simulator(tp.dyn_sys.f_num_simulation, tp.b, tp.dyn_sys.xa, x_col_fnc=None,
u_col_fnc=su1.f)
tt, xx, uu = sim.simulate()
return tt, xx, uu
def mab_so_gradient(designpoints,
step_size,
nr_samples=10000,
idx=0,
func=max):
""" Single objective gradient bandits algorithm.
Will first explore all design points once and then only take Upper-Confidence-Bound actions.
Since the output of the simulator will have multiple values (multi-objective), the idx
variable can be used to specify which value should be used by the MAB (default=TTF).
:param designpoints: [DesignPoint object] - List of DesignPoint objects (the candidates).
:param step_size: step size of the algorithm
:param nr_samples: number of samples
:param idx: index of simulator return value to use as objective
:param func: function to select the best DesignPoint (should be max or min).
:return: [(mean of samples, nr_samples)] - will return the mean of the sampled values
and the amount of samples taken for this dp.
"""
simulators = [Simulator(dp) for dp in designpoints]
k = len(designpoints) # total amount of bandits
H = np.zeros(k) # H from formula
P = np.ones(k) / k # probability per index
N = np.zeros(k) # amount of samples per index?
qt = np.zeros(k)
avg_reward = 0
for t in range(1, nr_samples + 1):
A = np.random.choice(k, 1, p=P)[0]
N[A] += 1
R = simulators[A].run()[idx]
qt[A] += (R - qt[A]) / N[A]
R /= 100000
avg_reward += (R - avg_reward) / t
baseline = avg_reward
H[A] += step_size * (R - baseline) * (1 - P[A])
for a in range(k):
if a != A:
H[a] -= step_size * (R - baseline) * P[a]
aux_exp = np.exp(H)
P = aux_exp / np.sum(aux_exp)
print(
"Best Gradient candidates:",
sorted(
set([
i for (i, val) in sorted(zip(np.arange(qt.size), qt),
key=lambda x: x[1],
reverse=True)
][:TOP_CANDIDATES])))
return list(zip(qt, N))
def simulate(self):
"""
This method is used to solve the resulting initial value problem
after the computation of a solution for the input trajectories.
"""
self.log_debug("Solving Initial Value Problem")
# calulate simulation time
T = self.dyn_sys.b - self.dyn_sys.a
##:ck: obsolete comment?
# Todo T = par[0] * T
# get list of start values
start = self.dyn_sys.xa
ff = self.dyn_sys.f_num_simulation
par = self.get_par_values()
# create simulation object
x_fncs, xdot_fncs, u_fnc = self.get_constrained_spline_fncs()
mpc_flag = self.nIt >= self.mpc_sim_threshold
self.simulator = Simulator(ff,
T,
start,
x_col_fnc=x_fncs,
u_col_fnc=u_fnc,
z_par=par,
dt=self._parameters['dt_sim'],
mpc_flag=mpc_flag)
self.log_debug("start: %s" % str(start))
# forward simulation
self.sim_data = self.simulator.simulate()
##:: S.simulate() is a method,
# returns a list [np.array(self.t), np.array(self.xt), np.array(self.ut)]
# self.sim_data is a `self.variable?` (initialized with None in __init__(...))
# convenient access
self.sim_data_tt, self.sim_data_xx, self.sim_data_uu = self.sim_data
def mab_so_gape_v(designpoints, a, b, m, nr_samples=1000, idx=0):
""" Single-objective MAB Gap-based Exploration with Variance (GapE-V).
:param designpoints: [DesignPoint object] - List of DesignPoint objects (the candidates).
:param a: degree of exploration
:param b: float - maximum expected value from samples
:param m: amount of designs to select
:param nr_samples: number of samples
:param idx: index of simulator return value to use as objective
:return: [(mean of samples, nr_samples)] - will return the mean of the sampled values
and the amount of samples taken for this dp.
"""
simulators = [Simulator(dp) for dp in designpoints]
ui = [(i, simulators[i].run()[idx])
for i in range(len(simulators))] # empirical means
oi = [0 for _ in range(len(simulators))]
T = [1 for _ in range(len(designpoints))]
gap_d = [0 for _ in range(len(designpoints))]
indices = [0 for _ in range(len(designpoints))]
for t in range(len(simulators), nr_samples):
sorted_indices = [
i for (i, val) in sorted(ui, key=lambda x: x[1], reverse=True)
]
i_star_up = sorted_indices[m]
i_star_down = sorted_indices[m + 1]
for i in sorted_indices[:m]:
gap_d[i] = ui[i][1] - ui[i_star_down][1]
for i in sorted_indices[m:]:
gap_d[i] = ui[i_star_up][1] - ui[i][1]
for i in sorted_indices:
indices[i] = -gap_d[i] + math.sqrt(
(2 * a * oi[i]) / T[i]) + (7 * a * b) / (3 * T[i])
j = indices.index(max(indices))
new_sample = simulators[j].run()[idx]
prev_ui = ui[j][1] # stores the current empiric mean
ui[j] = (j, ui[j][1] + (new_sample - ui[j][1]) / T[j])
# iterative variance calculation for o_i
oi[j] += ((new_sample - prev_ui) *
(new_sample - ui[j][1]) - oi[j]) / T[j]
T[j] += 1
print(
"Best GapE-V candidates:",
sorted(
set([
i for (i, val) in sorted(ui, key=lambda x: x[1], reverse=True)
][:TOP_CANDIDATES])))
return list(zip([x[1] for x in ui], T))
def simulate_naive_policy(n_trials, s_g, modality, usr):
planner = NaivePolicy()
model_path = "/home/yuhang/Documents/proactive_guidance/training_data/user" + str(usr)
sim = Simulator(planner, model_path)
traj_list = []
for i in range(n_trials):
traj_list.append(sim.run_trial((-1.0, 2.0, 0.0), s_g, modality, 20.0, tol=0.5))
fig, axes = plt.subplots()
for i in range(n_trials):
t, traj = traj_list[i]
axes.plot(traj[:, 0], traj[:, 1])
axes.axis("equal")
axes.scatter(s_g[0], s_g[1])
plt.show()
def sSAR(individuals, p, S, n):
""" Scalarized SAR implementation as presented in algorithm 3 of [Drugan&Nowe2014]
:param individuals: list of design points
:param p: number of individuals to select
:param S: list of scalarized functions
:param n: total number of samples
:return: Set of selected candidates
"""
accepted_arms = [set() for _ in range(len(S))]
K = len(individuals) # Number of rounds
A_all = [[i for i in range(K)] for _ in range(len(S))
] # Contains bandits for each scalarization function
A = [i for i in range(K)] # Total active arms
P_i = [p for _ in range(len(S))]
sims = [Simulator(i) for i in individuals] # Simulators per individual
N = [0 for _ in range(K)] # Number of samples per individual
ui = [[0 for _ in range(NR_OBJECTIVES)]
for _ in range(K)] # empirical reward vector
n_k = 0
LOG_K = 1 / 2 + sum([1 / i for i in range(2, K + 1)])
for k in range(1, K):
n_k_prev = n_k
n_k = math.ceil((1 / LOG_K) * ((n - K) / (K + 1 - k)))
samples = int(
n_k - n_k_prev) # Number of samples per individual in this phase
for i in A: # for all active bandits
for _ in range(
samples): # Sample each bandit and update empirical vector
N[i] += 1
reward_vector = sims[i].run()
ui[i] = update_empirical_mean(normalize(reward_vector), ui[i],
N[i])
for i in range(len(S)):
max_gap_idx, accepted = delta_pk_ij(ui, A_all[i], S[i],
P_i[i] - len(accepted_arms[i]))
A_all[i].remove(max_gap_idx)
if accepted: # Store the arms that are accepted by a function for this round
accepted_arms[i].add(max_gap_idx)
# Updates A to only include any arm that has not yet been removed by an F_j
A = set().union(*A_all)
# invert normalization of mean values
ui = [normalize(i, invert=True) for i in ui]
return set.union(*accepted_arms), ui, N
def validate_free_space_policy(planner, s_g, modality, path, model_path):
fig, axes = plt.subplots()
planner.visualize_policy(axes)
fig.savefig(path + "/value_func.png")
sim = Simulator(planner, model_path)
n_trials = 30
traj_list = []
for i in range(n_trials):
traj_list.append(sim.run_trial((-1.0, 2.0, 0.0), s_g, modality, 30.0, tol=0.5))
fig, axes = plt.subplots()
for i in range(n_trials):
t, traj = traj_list[i]
axes.plot(traj[:, 0], traj[:, 1])
axes.axis("equal")
axes.scatter(s_g[0], s_g[1])
fig.savefig(path + "/simulation.png")
def simulate_naive_policy(n_trials, s_g, modality, usr):
planner = NaivePolicy()
model_path = "/home/yuhang/Documents/proactive_guidance/training_data/user" + str(
usr)
sim = Simulator(planner, model_path)
traj_list = []
for i in range(n_trials):
traj_list.append(
sim.run_trial((-1.0, 2.0, 0.0), s_g, modality, 20.0, tol=0.5))
fig, axes = plt.subplots()
for i in range(n_trials):
t, traj = traj_list[i]
axes.plot(traj[:, 0], traj[:, 1])
axes.axis("equal")
axes.scatter(s_g[0], s_g[1])
plt.show()
def validate_free_space_policy(planner, s_g, modality, path, model_path):
fig, axes = plt.subplots()
planner.visualize_policy(axes)
fig.savefig(path + "/value_func.png")
sim = Simulator(planner, model_path)
n_trials = 30
traj_list = []
for i in range(n_trials):
traj_list.append(
sim.run_trial((-1.0, 2.0, 0.0), s_g, modality, 30.0, tol=0.5))
fig, axes = plt.subplots()
for i in range(n_trials):
t, traj = traj_list[i]
axes.plot(traj[:, 0], traj[:, 1])
axes.axis("equal")
axes.scatter(s_g[0], s_g[1])
fig.savefig(path + "/simulation.png")
def esSR(individuals, S, n):
""" Efficient Scalarized Succesive Reject multi-armed bandit algorithm.
Algorithm2 as presented in [Drugan & Nowe (2014)].
Samples a list of individual design points in phases and will reject a design point
per phase with multiple scalarization functions (S).
At the end only one design point will remain, which is considered the best arm.
:param individuals: list of design points
:param S: list of scalarized functions
:param n: total number of samples
:return:
"""
K = len(individuals)
A = [i for i in range(K)] # Contains all active bandits
A_all = [[i for i in range(K)] for _ in range(len(S))
] # Contains bandits for each scalarization function
N = [0 for _ in range(K)] # Number of samples per bandit
ui = [[0 for _ in range(NR_OBJECTIVES)]
for _ in range(K)] # empirical reward vector
sims = [Simulator(i) for i in individuals]
n_k = 0
LOG_K = 1 / 2 + sum([1 / i for i in range(2, K + 1)])
for k in range(1, K):
n_k_prev = n_k
n_k = math.ceil(1 / LOG_K * (n - K) / (K + 1 - k))
samples = int(n_k - n_k_prev) # Number of samples per phase
for i in A: # for all active bandits
for _ in range(
samples): # Sample each bandit and update empirical vector
N[i] += 1
ui[i] = update_empirical_mean(sims[i].run(), ui[i], N[i])
for i in range(
len(S)): # for each scalarization function dismiss a bandit
A_j = A_all[i]
scalarized_rewards = [(z, S[i](ui[z])) for z in A_j if z in A_j]
idx = min(scalarized_rewards, key=lambda t: t[1])[0]
A_j.remove(
idx) # deletes the worst individual of this scalarization func
A = set().union(
*A_all
) # Updates A to only include any arm that has not yet been removed by an F_j
print("Accepted:", A)
return list(zip(ui, N))
def SAR(individuals, m, nr_samples=1000):
""" Scalarized multi-objective MAB evaluation via Successive Accept Reject (SAR).
:param individuals: individuals provided by the ga (must have fitness attributes)
:param m: int - length of best-arm set
:param nr_samples: int - amount of samples
:return: [(idx, sampled_value)]
"""
simulators = [Simulator(d) for d in individuals]
D = len(individuals)
A = [i for i in range(D)]
N = [0 for _ in range(D)]
ui = [(i, 0) for i in range(len(simulators))] # empirical means
S = set()
LOG_D = 1 / 2 + sum([1 / i for i in range(2, D + 1)])
m_o = m
for k in range(1, D):
samples = int(f_n_k(k, nr_samples, D) - f_n_k(k - 1, nr_samples, D))
for i in A:
for _ in range(samples):
new_sample = l_scale(list(simulators[i].run()) +
[individuals[i].evaluate_size()],
weights=(1, -1, -1))
N[i] += 1
ui[i] = (i, ui[i][1] + (new_sample - ui[i][1]) / N[i])
sorted_indices = [
i for (i, val) in sorted(ui, key=lambda x: x[1], reverse=True)
if i in A
]
i_star_up = sorted_indices[m_o]
i_star_down = sorted_indices[m_o + 1]
gap_d = [0 for _ in range(D)]
for i in sorted_indices[:m_o]:
gap_d[i] = ui[i][1] - ui[i_star_down][1]
for i in sorted_indices[m_o:]:
gap_d[i] = ui[i_star_up][1] - ui[i][1]
j = gap_d.index(max(gap_d))
if j == sorted_indices[0]:
S.add(j)
m_o -= 1
A = [i for i in A if i != j]
return ui
def pareto_ucb1(individuals, k, nr_samples=500):
""" Implementation of the pareto Upper-Confidence-Bound1 (pareto UCB1) pseudocode [Drugan&Nowe(2013)]
:param individuals: individuals provided by the ga (must have fitness attributes)
:param k: The number of individuals to select.
:return: [(mean of samples, nr_samples)] - will return the mean of the sampled values
and the amount of samples taken for this dp.
:return: [(idx, sampled_value)]
"""
n = len(individuals)
simulators = {individuals[i]: Simulator(individuals[i]) for i in range(n)}
# Samples per individual
N = {individuals[i]: 1 for i in range(n)}
# Empirical mean vector per individual
ui = {
individuals[i]: list(normalize(simulators[individuals[i]].run()))
for i in range(n)
}
# individual fitness values are empirical means
for i in individuals:
mttf, pow_usage, size = ui[i]
i.fitness.values = (mttf, pow_usage, size)
samples = len(individuals)
while samples < nr_samples:
A_star = sortNondominated(individuals, k, first_front_only=True)[0]
# Adds confidence interval
add_confidence_interval(individuals, list(N.values()), len(A_star))
A_p = sortNondominated(individuals, k, first_front_only=True)[0]
# Removes confidence interval
add_confidence_interval(individuals,
list(N.values()),
len(A_star),
subtract=True)
a = np.random.choice(A_p)
N[a] += 1
samples += 1
ui[a] = update_empirical_mean(normalize(simulators[a].run()), ui[a],
N[a])
a.fitness.values = ui[a]
return [normalize(ui[individuals[i]], invert=True)
for i in range(n)], [N[individuals[i]] for i in range(n)]
def get_simulator_example(cap1=100,
cap2=100,
loc1=(0, 0),
loc2=(1, 1),
app1=50,
app2=50):
c1 = Component(cap1, loc1)
c2 = Component(cap2, loc2)
a1 = Application(app1)
a2 = Application(app2)
dp = DesignPoint([c1, c2], [a1, a2], [(c1, a1), (c2, a2)])
return Simulator(dp)
def evaluate_timing(timing):
traffic_light = PhaseModifier("node1")
full_timings = [3] * 8
full_timings[0] = timing[0]
full_timings[4] = timing[1]
controller = StaticTrafficLightController(controller=traffic_light,
sequence=list(range(8)),
timings=full_timings)
sim = Simulator()
sim.add_simulation_component(SimulationOutputParser)
sim.add_tickable(controller)
if not sim.run(sumocfg1, time_steps=2000, gui=False):
return sim.results
return False
def mab_so_epsilon_greedy(designpoints, e, nr_samples=10000, idx=0, func=max):
""" Single objective epsilon-greedy evaluation of the given list of design points.
Will first explore all design points once and then only take epsilon-greedy actions.
Since the output of the simulator will have multiple values (multi-objective), the idx
variable can be used to specify which value should be used by the MAB (default=TTF).
:param designpoints: [DesignPoint object] - List of DesignPoint objects (the candidates).
:param e: epsilon value between [0.0 - 1.0] indicating the probability to explore a random dp.
:param nr_samples: number of samples
:param idx: index of simulator return value to use as objective
:param func: function to select the best DesignPoint (should be max or min).
:return: [(mean of samples, nr_samples)] - will return the mean of the sampled values
and the amount of samples taken for this dp.
"""
simulators = [Simulator(dp) for dp in designpoints]
samples = [1 for _ in simulators]
qt = [sim.run()[idx]
for sim in simulators] # Q_t = est value of action a at timestep t
for _ in range(len(simulators), nr_samples):
if random.random() < e: # epsilon exploration
a = random.randint(0, len(simulators) - 1)
else: # greedy exploitation
a = random.choice(
[i for i, val in enumerate(qt) if val == func(qt)])
new_sample = simulators[a].run()[idx]
samples[a] += 1
# Incremental implementation of MAB [Sutton & Barto(2011)].
qt[a] += (new_sample - qt[a]) / samples[a]
print(
"Best e-greedy candidates:",
sorted(
set([
i for (i, val) in sorted(zip(np.arange(len(qt)), qt),
key=lambda x: x[1],
reverse=True)
][:TOP_CANDIDATES])))
return list(zip(qt, samples))
def simulator(request):
from main import AutotestingApp
application = AutotestingApp()
simulator = Simulator(application)
def fin():
simulator.clean_queue()
from kivy.lang import Builder
files = list(Builder.files)
for filename in files:
Builder.unload_file(filename)
kivy_style_filename = os.path.join(kivy_data_dir, 'style.kv')
if not kivy_style_filename in Builder.files:
Builder.load_file(kivy_style_filename, rulesonly=True)
request.addfinalizer(fin)
return simulator
def mab_so_ucb(designpoints, c, nr_samples=10000, idx=0, func=max):
""" Single objective Upper-Confidence-Bound (UCB) action selection
Will first explore all design points once and then only take Upper-Confidence-Bound actions.
Since the output of the simulator will have multiple values (multi-objective), the idx
variable can be used to specify which value should be used by the MAB (default=TTF).
:param designpoints: [DesignPoint object] - List of DesignPoint objects (the candidates).
:param c: degree of exploration
:param nr_samples: number of samples
:param idx: index of simulator return value to use as objective
:param func: function to select the best DesignPoint (should be max or min).
:return: [(mean of samples, nr_samples)] - will return the mean of the sampled values
and the amount of samples taken for this dp.
"""
simulators = [Simulator(dp) for dp in designpoints]
samples = [1 for _ in simulators]
qt = [sim.run()[idx]
for sim in simulators] # Q_t = est value of action a at timestep t
for t in range(len(simulators), nr_samples):
actions = [
qt[i] + c * math.sqrt(math.log(t) / samples[i])
for i in range(len(simulators))
]
a = random.choice(
[i for i, val in enumerate(actions) if val == func(actions)])
samples[a] += 1
new_sample = simulators[a].run()[idx]
qt[a] += (new_sample - qt[a]) / samples[a]
print(
"Best UCB candidates:",
sorted(
set([
i for (i, val) in sorted(zip(np.arange(len(qt)), qt),
key=lambda x: x[1],
reverse=True)
][:TOP_CANDIDATES])))
return list(zip(qt, samples))
def compareChurnStable():
s = Simulator()
random.seed(125)
s.setupSimulation(strategy= "churn", workMeasurement= "one", numNodes= 1000, numTasks = 100000, churnRate =0)
loads1, medians1, means1, maxs1, devs1 = s.simulateLoad()
random.seed(125)
s=Simulator()
s.setupSimulation(strategy= "churn", workMeasurement= "one", numNodes= 1000, numTasks = 100000, churnRate =0.1)
loads2 = s.simulateLoad()[0]
for i in range(0,len(loads1), 5):
x1= loads1[i]
x2 =loads2[i]
colors = ["k", "w"]
labels = ["No Strategy","Churn"]
plt.hist([x1,x2], 25, normed =1, color=colors, label=labels)
plt.legend(loc=0)
plt.title('Churn vs No Strategy at Tick ' + str(i))
plt.xlabel('Tasks Per Node')
plt.ylabel('Fraction of the Network')
#plt.ylim(0, 0.05)
plt.show()
from network import Network
from process import Process
from simulation import Simulator
from numpy import inf
import oracle
import random
import action
import time
fault_rates = [100]
start = time.clock()
sim = Simulator()
sim.run_fault_experiment(fault_rates, 1, 100)
end = time.clock()
print "Experiment finished! (", end - start, "s)"
sim.record_results("Linear-results-faults_1-10.txt", fault_rates)
def testChurnSteps():
s = Simulator()
s.setupSimulation(strategy="churn", workMeasurement="one", numNodes=1000, numTask=100000, churnRate=0.001)
loads, medians, means, maxs, devs = s.simulateLoad()
print(medians)
print(devs)
