Twiddle - Artificial Intelligence for Robotics Udacity English Japanese

Twiddle - Artificial Intelligence for RoboticsYouTube↗ IA↗(long shot)
Education 13.8K views 64 likes 2 dislikes 18 Mar 2012

Udacity 651K subscribers United States YouTube↗

Click to Play Video

In Twiddle, we're trying to optimize for a set of parameters.

いじり回しではパラメータの集合を最適化します

To do so, our function run() must return a goodness.

そのためにはrun関数でいい値を返す必要があります

This goodness value might be the average crosstrack error.

このいい値はクロストラックエラーの平均値でいいでしょう

Say I wanted to implement Twiddle so as to minimize the average crosstrack error.

クロストラックエラーの平均を最小化するよういじり回しを実装したいと思います

If that's the case, then the output of run depends on the three parameters.

その状況ではrunの出力は 3つのパラメータに依存します

Here's how Twiddle works.

いじり回しは次のように動作します

Build a parameter vector of our 3-target parameters, and initialize it with zero.

3つの目標パラメータのベクトルを作り 0で初期化します

Also, build a vector of potential changes that you want to probe and initialize them for now with 1.

調査したい潜在的変化のベクトルdpも作りまずは1で初期化します

Then you can run our command run( ) with our parameters,

これらのパラメータでコマンドrunを実行できます

and whatever it outputs is our best error so far.

出力はすべて一番いい誤差best＿errとします

Now we wish to modify p as to make the error smaller.

誤差をより小さくするようpを修正しましょう

That's where Twiddle comes in. It's a really smart algorithm, I believe.

ここでいじり回しが登場しますこれは本当に賢いアルゴリズムです

We sequentially go through these parameters.

これらのパラメータを連続的に調べます

Obviously, you shouldn't write 3. You should write len of p.

ここは当然3と書くべきではありません pの長さを書くべきです

First we tried to increase p by our probing value,

まず調査の値によりpを増やすことを試みます

compute a new error for this new modified p.

修正したpに対する新たな誤差errを計算します

If this new error is better than our best error, then we do two things.

新たなerrがこれまでのbest＿errよりいい場合2つのことを行います

First, we set best_err to err, and we even modify our dp to a slightly larger value

まずbest＿errに新たなerrを設定します次にdpをわずかに大きな値へ修正します

by multiplying it with 1.1.

そのためにdpへ1．1を掛けます

Otherwise, we try the other way.

値が悪い場合他の方法を試みます

We subtract dp from p--and we have to do it twice now because we added it before.

pからdpを減算します　以前一度加算しているので減算を2回行う必要があります

Then we do the same thing again as over here. I'm not going to write it out.

すべて書きませんが同様の処理を再び行います

We check whether the error is better than our best error, we retain it, and we multiply dp by 1.1.

errがbest＿errよりいいかチェックしそれを保持してdpに1．1を掛けます

But if both of those fail--this one over here and this one over here--

しかしこれらのいずれにも失敗した場合

we set p[ i ] back to the original value, and we decrease our probing thing over here,

p［i］に元の値を設定し調査の値を減らします

say, by multiplying it with 0.9.

そのためにdpへ0．9を掛けます

That's the core of Twiddle, and what it really does is for each coordinate in isolation

これがいじり回しの核心です各座標に対し個別に実際に行うのです

it moves our parameter down a little bit by this value over here.

このdpの値でパラメータを少し上下させます

If it then finds a better solution, it retains it, and it even increments the probing interval.

よりよい解を見つけたらそれを保持し調査の間隔を増やします

If it fails to find a better solution, it goes back to the original and decreases our probing interval.

いい解を見つけることに失敗した場合元の値に戻し調査の間隔を減らします

We do this entire thing so long as the sum of the dp's is larger than the threshold.

このすべての処理をdpの総和が閾値より大きい間行います

Somewhere in here we say while some of dp is larger than 0.00001.

この辺りでdpの総和が 0．00001より大きいとしましょう

It's hard to read, but I hope you can follow it.

読みにくいですが理解できると思います

This is Twiddle. Let me put this into pictures.

これがいじり回しですこれを図で描いてみましょう

We have three parameters--0, 0, 0.

3つのパラメータ0、0、0があります

Then in the first iteration, we bump one of the parameters up and see if it improves the error.

最初のループでパラメータの1つを増やしますさらに誤差を改善するかも調べます

If that's the case, we retain it. Then we go to the second parameter. We bump it up. It might not work.

改善されれば値を保持し 2つ目のパラメータへ移りそれを増やします

We bump it down and maybe retain that one, and so on.

ダメなら減らして保持するといった感じです

Now, as we keep bumping up, we might find that neither bumping up nor bumping down helps.

増やし続けると増加も減少も効果がないと分かるかもしれません

What we do instead is we retain the original solution but make our probing interval

増減の代わりに元の解を保持しますが

smaller than before by a factor of 0.9.

調査の間隔を係数0．9で以前より小さくします

In doing so, we can zoom in more and more into a detailed parameter until it finally converges.

そうすることで最終的に収束するまで詳細パラメータを突き詰めていきます

It's local hill climber, but it happens to be really, really efficient.

局所的な山登り法ですが極めて効果的です

Similar videos:

Separated Gaussians 2 - Artificial Intelligence for Robotics
Udacity
14.3K views 9 likes 0 dislikes 29 May 2012100% match

Expansion Grid - Artificial Intelligence for Robotics
Udacity
10.4K views 2 likes 0 dislikes 29 May 2012100% match

Exact Motion - Artificial Intelligence for Robotics
Udacity
15.1K views 12 likes 1 dislikes 19 Feb 2012100% match

Programming Exercise - Artificial Intelligence for Robotics
Udacity
35.6K views 144 likes 6 dislikes 23 Feb 2015100% match

Probability - Artificial Intelligence for Robotics
Udacity
7.6K views 6 likes 2 dislikes 02 Mar 2012100% match

Circular Motion - Artificial Intelligence for Robotics
Udacity
17.2K views 19 likes 2 dislikes 23 Feb 2015100% match

Shifting the Mean - Artificial Intelligence for Robotics
Udacity
15.9K views 7 likes 2 dislikes 29 May 2012100% match

Inexact Motion 3 - Artificial Intelligence for Robotics
Udacity
15K views 6 likes 1 dislikes 19 Feb 2012100% match

Bayes' Rule - Artificial Intelligence for Robotics
Udacity
12.7K views 8 likes 4 dislikes 02 Mar 2012100% match

Parameter Update - Artificial Intelligence for Robotics
Udacity
28.7K views 5 likes 3 dislikes 29 May 2012100% match

Kalman Prediction - Artificial Intelligence for Robotics
Udacity
10.5K views 6 likes 1 dislikes 29 May 2012100% match

Predict Function - Artificial Intelligence for Robotics
Udacity
15.2K views 4 likes 0 dislikes 29 May 2012100% match

Generalized Uniform Distribution - Artificial Intelligence for Robotics
Udacity
34.9K views 12 likes 0 dislikes 19 Feb 2012100% match

State Space - Artificial Intelligence for Robotics
Udacity
10.9K views 7 likes 1 dislikes 29 May 2012100% match

Normalized Sense Function - Artificial Intelligence for Robotics
Udacity
17.7K views 8 likes 0 dislikes 19 Feb 2012100% match

PD Controller - Artificial Intelligence for Robotics
Udacity
22.3K views 21 likes 1 dislikes 30 May 2012100% match

Sensing - Artificial Intelligence for Robotics
Udacity
6.4K views 1 likes 0 dislikes 23 Feb 2015100% match

Planning Solution - Artificial Intelligence for Robotics
Udacity
620 views 0 likes 0 dislikes 16 Mar 2012100% match

Implement P Controller - Artificial Intelligence for Robotics
Udacity
13.7K views 7 likes 0 dislikes 30 May 2012100% match

Formal Definition of Probability 2 - Artificial Intelligence for Robotics
Udacity
11.3K views 7 likes 0 dislikes 19 Feb 2012100% match

Separated Gaussians - Artificial Intelligence for Robotics
Udacity
22.9K views 13 likes 0 dislikes 29 May 2012100% match

A* in Action - Artificial Intelligence for Robotics
Udacity
21.4K views 236 likes 2 dislikes 30 May 2012100% match

Moving Robot - Artificial Intelligence for Robotics
Udacity
22.1K views 14 likes 0 dislikes 29 May 2012100% match

Normalize Distribution - Artificial Intelligence for Robotics
Udacity
20.5K views 11 likes 4 dislikes 19 Feb 2012100% match

Path Smoothing - Artificial Intelligence for Robotics
Udacity
18.2K views 35 likes 0 dislikes 30 May 2012100% match

Move Function - Artificial Intelligence for Robotics
Udacity
11.5K views 15 likes 1 dislikes 19 Feb 2012100% match

Move 1000 - Artificial Intelligence for Robotics
Udacity
11.6K views 4 likes 0 dislikes 19 Feb 2012100% match

Gaussian Motion - Artificial Intelligence for Robotics
Udacity
15.5K views 8 likes 0 dislikes 29 May 2012100% match

Sum of Probabilities - Artificial Intelligence for Robotics
Udacity
18.2K views 6 likes 0 dislikes 19 Feb 2012100% match

Predicting the Peak - Artificial Intelligence for Robotics
Udacity
15.6K views 4 likes 2 dislikes 29 May 2012100% match

Settings Privacy About Contact Donate