この jupyter notebook ファイルは ISSP Data Repository (v3.0.0 branch) から入手できます。

インタラクティブに実行する

以下の流れで、PHYSBO をインタラクティブに実行することができます。

PHYSBO から次に実行するパラメータを得ます。
PHYSBO の外部で評価値を得ます。
評価値をPHYSBOに登録します。

例えば、以下の様な場合に適しています。

人手による実験を行い、その評価値をPHYSBOに与えたい。
simulator の実行を別プロセスで行うなど、柔軟に実行制御を行いたい。

探索候補データの準備

これまでのチュートリアルと同様、データセットファイル s5-210.csv を保存し、次のように読み出します。

[1]:

import physbo

import numpy as np


def load_data():
    A =  np.asarray(np.loadtxt('s5-210.csv',skiprows=1, delimiter=',') )
    X = A[:,0:3]
    t  = -A[:,3]
    return X, t

X, t = load_data()
X = physbo.misc.centering(X)

Cythonized version of physbo is used

simulator の定義

[2]:

class Simulator:
    def __init__( self ):
        _, self.t = load_data()

    def __call__( self, action ):
        return self.t[action]

最適化の実行

[3]:

# policy のセット
policy = physbo.search.discrete.Policy(test_X=X)

# シード値のセット
policy.set_seed( 0 )

各探索ステップでは以下の処理を行っています。

max_num_probes=1, simulator=None として random_search または bayes_search を実行して action ID (パラメータ)を得る。
t = simulator(actions) により評価値(の array) を得る。
policy.write(actions, t)により action ID (パラメータ) に対する評価値を登録する。
physbo.search.utility.show_search_results により履歴を表示する。

以下では、ランダムサンプリングを２回（1st, and 2nd steps）、ベイズ最適化による提案を２回（3rd, and 4th steps）を行います。

[4]:

simulator = Simulator()

''' 1st step (random sampling) '''
actions = policy.random_search(max_num_probes=1, simulator=None)
t  = simulator(actions)
policy.write(actions, t)
physbo.search.utility.show_search_results(policy.history, 10)

''' 2nd step (random sampling) '''
actions = policy.random_search(max_num_probes=1, simulator=None)
t = simulator(actions)
policy.write(actions, t)
physbo.search.utility.show_search_results(policy.history, 10)

''' 3rd step (bayesian optimization) '''
actions = policy.bayes_search(max_num_probes=1, simulator=None, score='EI', interval=0,  num_rand_basis = 5000)
t = simulator(actions)
policy.write(actions, t)
physbo.search.utility.show_search_results(policy.history, 10)

''' 4-th step (bayesian optimization) '''
actions = policy.bayes_search(max_num_probes=1, simulator=None, score='EI', interval=0,  num_rand_basis = 5000)
t = simulator(actions)
policy.write(actions, t)
physbo.search.utility.show_search_results(policy.history, 10)

interactive mode starts ...

current best f(x) = -1.070602 (best action = 15673)
list of simulation results
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=-1.070602 (action = 15673)


current best f(x) = -1.070602 (best action = 15673)
list of simulation results
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=-1.070602 (action = 15673)
f(x)=-1.153410 (action = 16489)


Start the initial hyper parameter searching ...
Done

Start the hyper parameter learning ...
0 -th epoch marginal likelihood -3.530887145012235
50 -th epoch marginal likelihood -3.530886068830362
100 -th epoch marginal likelihood -3.530887353880807
150 -th epoch marginal likelihood -3.5308873663578604
200 -th epoch marginal likelihood -3.530887366409382
250 -th epoch marginal likelihood -3.5308873664094484
300 -th epoch marginal likelihood -3.5308873664094484
350 -th epoch marginal likelihood -3.530887366409448
400 -th epoch marginal likelihood -3.530887366409449
450 -th epoch marginal likelihood -3.530887366409449
500 -th epoch marginal likelihood -3.530887366409448
Done

current best f(x) = -0.993712 (best action = 11874)
list of simulation results
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=-1.070602 (action = 15673)
f(x)=-1.153410 (action = 16489)
f(x)=-0.993712 (action = 11874)


Start the initial hyper parameter searching ...
Done

Start the hyper parameter learning ...
0 -th epoch marginal likelihood -3.9335555735844907
50 -th epoch marginal likelihood -3.933559178581591
100 -th epoch marginal likelihood -3.9335601169323873
150 -th epoch marginal likelihood -3.933560116971464
200 -th epoch marginal likelihood -3.9335601170149546
250 -th epoch marginal likelihood -3.933560117015064
300 -th epoch marginal likelihood -3.933560117015065
350 -th epoch marginal likelihood -3.933560117015066
400 -th epoch marginal likelihood -3.933560117015066
450 -th epoch marginal likelihood -3.933560117015066
500 -th epoch marginal likelihood -3.933560117015065
Done

current best f(x) = -0.985424 (best action = 8061)
list of simulation results
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=-1.070602 (action = 15673)
f(x)=-1.153410 (action = 16489)
f(x)=-0.993712 (action = 11874)
f(x)=-0.985424 (action = 8061)

中断と再開

以下の predictor, training, history を外部ファイルに保存することで、最適化プロセスを中断し、途中から再開することができます。

predictor: 目的関数の予測モデル
training: predictor の学習に用いるデータ (physbo.Variable オブジェクト）
history: 最適化実行の履歴 (physbo.search.discrete.History オブジェクト)

[5]:

policy.save(file_history='history.npz', file_training='training.npz', file_predictor='predictor.dump')

[6]:

# policy を削除
del policy

# 保存した policy をロード
policy = physbo.search.discrete.Policy(test_X=X)
policy.load(file_history='history.npz', file_training='training.npz', file_predictor='predictor.dump')

''' 5-th step (bayesian optimization) '''
actions = policy.bayes_search(max_num_probes=1, simulator=None, score='EI', interval=0,  num_rand_basis = 5000)
t = simulator(actions)
policy.write(actions, t)
physbo.search.utility.show_search_results(policy.history, 10)

# predictor と training を個別に指定することも可
''' 6-th step (bayesian optimization) '''
actions = policy.bayes_search(max_num_probes=1,
                                            predictor=policy.predictor, training=policy.training,
                                            simulator=None, score='EI', interval=0,  num_rand_basis = 5000)
t = simulator(actions)
policy.write(actions, t)
physbo.search.utility.show_search_results(policy.history, 10)

Start the initial hyper parameter searching ...
Done

Start the hyper parameter learning ...
0 -th epoch marginal likelihood -5.282317837949537
50 -th epoch marginal likelihood -5.314394752726803
100 -th epoch marginal likelihood -5.333676604278033
150 -th epoch marginal likelihood -5.344148457109117
200 -th epoch marginal likelihood -5.349388232794448
250 -th epoch marginal likelihood -5.351839835112543
300 -th epoch marginal likelihood -5.352933499846448
350 -th epoch marginal likelihood -5.353415283021385
400 -th epoch marginal likelihood -5.3536398281059405
450 -th epoch marginal likelihood -5.3537614274261145
500 -th epoch marginal likelihood -5.353841697978967
Done

current best f(x) = -0.985424 (best action = 8061)
list of simulation results
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=-1.070602 (action = 15673)
f(x)=-1.153410 (action = 16489)
f(x)=-0.993712 (action = 11874)
f(x)=-0.985424 (action = 8061)
f(x)=-1.033129 (action = 8509)


Start the initial hyper parameter searching ...
Done

Start the hyper parameter learning ...
0 -th epoch marginal likelihood -7.1246867953581425
50 -th epoch marginal likelihood -7.125240732973938
100 -th epoch marginal likelihood -7.125738402977026
150 -th epoch marginal likelihood -7.126320570892344
200 -th epoch marginal likelihood -7.127007900107863
250 -th epoch marginal likelihood -7.127824732094982
300 -th epoch marginal likelihood -7.1288019858435145
350 -th epoch marginal likelihood -7.129979143770724
400 -th epoch marginal likelihood -7.131406793529879
450 -th epoch marginal likelihood -7.133149836164181
500 -th epoch marginal likelihood -7.135291360142487
Done

current best f(x) = -0.985424 (best action = 8061)
list of simulation results
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=0.000000 (action = 0)
f(x)=-1.070602 (action = 15673)
f(x)=-1.153410 (action = 16489)
f(x)=-0.993712 (action = 11874)
f(x)=-0.985424 (action = 8061)
f(x)=-1.033129 (action = 8509)
f(x)=-1.061677 (action = 11352)

[ ]: