人工知能学会 第25回知識流通ネットワーク研究会発表 http://sigksn.html.xdomain.jp/conf25/index.html
システム障害解析に関する専門家知識の抽出、グラフ化、DB化を行った際得られた知見と、知識流通手段としての知識グラフの可能性と課題を考察した結果を報告します。
Knowledge graphs have been getting attention because of its relevance to interpretable AI. Not only that, they also can be useful as a knowledge sharing mean which enable non-experts to utilize experts’ knowledge. We aim to report findings from constructing a knowledge graph through eliciting experts’ knowledge and building a knowledge database. We also suggest the possibilities and issues of knowledge graph as a knowledge sharing mean.
50. (参考)ベイジアンネットワークの概要(2)
解放量過大 断片化 GC長時間化
T F
T T 0.9 0.1
T F 0.7 0.3
F T 0.6 0.4
F F 0.1 0.9
解放量
過大
断片化
T F
T 0.8 0.2
F 0.1 0.9
解放量過大
T F
0.4 0.6
③GC処理が長
時間化
①メモリ解放量
が過大
②メモリ断片化が発
生
0.4
0.6
0.38
0.62
0.57
T
F
T
F
T
F
𝑷 𝑪 𝑬 =
𝑷 𝑬 𝑪 𝑷 𝑪
Ʃ𝑷 𝑬 𝑪′
𝑷 𝑪′
0.43
観測値が何も無い
ときの確率を計算
する(事前確率)
50
51. (参考)ベイジアンネットワークの概要(3)
解放量過大 断片化 GC長時間化
T F
T T 0.9 0.1
T F 0.7 0.3
F T 0.6 0.4
F F 0.1 0.9
解放量
過大
断片化
T F
T 0.8 0.2
F 0.1 0.9
解放量過大
T F
0.4 0.6
③GC処理が長
時間化
①メモリ解放量
が過大
②メモリ断片化が発
生
0.84
1.0
0.0
T
F
T
F
T
F
(1) エビデンス
(観測値)の設定
0.87
(2) 確率伝搬
(2) 確率伝搬
0.16
0.13
観測値が得られた場合,
確率を再計算する
(事後確率)
51
Notas do Editor
Now we are going to see the conditional probabilities a little deeper.
In a Bayesian network, the conditional probabilities are in form of tables, which is called conditional probability table. We call it CPT as a short form.
In a Bayesian network, every node has a CPT, which contains probabilities of each condition of parent nodes.
Then, it calculate the prior probabilities.
With the formula on the screen, probability of a cause C when Effect E has given is calculated. We call it “Probability of C given E”.
(Probability of E given C by Probability of C over sum of all probabilities of E given every C by probability of every C.)
They are the hypothesis before taking any evidence into account.
And, then, you obtain an evidence, which indicate that the old generation of Java heap is fragmented, it is set to the model.
The probability of T on the node No.2 become ONE.point.ZERO.
The change propagate to the other nodes. As you see, the probabilities of T on the other nodes also become high.
In that manner, utilizing evidences, taking them into account, we can narrow down the true root cause.