ãã€ããŒãã©ã¡ãŒã¿èª¿æŽã®ããã®ãã€ãºæé©åã®å æ¬çãªã¬ã€ãããã®ååãå©ç¹ãå®è·µçãªå®è£ ãããã³é«åºŠãªãã¯ããã¯ãç¶²çŸ ããŸãã
ãã€ããŒãã©ã¡ãŒã¿èª¿æŽïŒãã€ãºæé©åã®ç¿åŸ
æ©æ¢°åŠç¿ã®åéã§ã¯ãã¢ãã«ã®ããã©ãŒãã³ã¹ã¯ããã®ãã€ããŒãã©ã¡ãŒã¿ã«ãã£ãŠå€§ãã圱é¿ãåããããšããããããŸãããã¬ãŒãã³ã°äžã«åŠç¿ãããã¢ãã«ãã©ã¡ãŒã¿ãšã¯ç°ãªãããã€ããŒãã©ã¡ãŒã¿ã¯ãã¬ãŒãã³ã°ããã»ã¹ãéå§ãããåã«èšå®ãããŸããæé©ãªãã€ããŒãã©ã¡ãŒã¿æ§æãèŠã€ããã®ã¯ãå°é£ã§æéã®ãããäœæ¥ã«ãªãå¯èœæ§ããããŸããããã§ããã€ããŒãã©ã¡ãŒã¿èª¿æŽãã¯ããã¯ãç»å Žããäžã§ããã€ãºæé©åã¯ã匷åã§å¹ççãªã¢ãããŒããšããŠéç«ã£ãŠããŸãããã®èšäºã§ã¯ããã€ãºæé©åã®å æ¬çãªã¬ã€ããæäŸãããã®ååãå©ç¹ãå®è·µçãªå®è£ ãããã³é«åºŠãªãã¯ããã¯ãç¶²çŸ ããŸãã
ãã€ããŒãã©ã¡ãŒã¿ãšã¯ïŒ
ãã€ããŒãã©ã¡ãŒã¿ã¯ããã¬ãŒãã³ã°ããã»ã¹äžã«ããŒã¿ããåŠç¿ãããªããã©ã¡ãŒã¿ã§ãããããã¯åŠç¿ããã»ã¹èªäœãå¶åŸ¡ããã¢ãã«ã®è€éããåŠç¿çãããã³å šäœçãªåäœã«åœ±é¿ãäžããŸãããã€ããŒãã©ã¡ãŒã¿ã®äŸãšããŠã¯ã次ã®ãã®ããããŸãã
- åŠç¿çïŒãã¥ãŒã©ã«ãããã¯ãŒã¯ã«ãããåŸé éäžäžã®ã¹ããããµã€ãºãå¶åŸ¡ããŸãã
- ã¬ã€ã€ãŒ/ãã¥ãŒãã³ã®æ°ïŒãã¥ãŒã©ã«ãããã¯ãŒã¯ã®ã¢ãŒããã¯ãã£ãå®çŸ©ããŸãã
- æ£ååã®åŒ·åºŠïŒéåŠç¿ãé²ãããã«ã¢ãã«ã®è€éããå¶åŸ¡ããŸãã
- ã«ãŒãã«ãã©ã¡ãŒã¿ïŒãµããŒããã¯ã¿ãŒãã·ã³ïŒSVMïŒã®ã«ãŒãã«é¢æ°ãå®çŸ©ããŸãã
- ããªãŒã®æ°ïŒã©ã³ãã ãã©ã¬ã¹ãã«ãããæ±ºå®æšã®æ°ã決å®ããŸãã
ãã€ããŒãã©ã¡ãŒã¿ã®é©åãªçµã¿åãããèŠã€ããããšã§ãã¢ãã«ã®ããã©ãŒãã³ã¹ãå€§å¹ ã«åäžãããããåªããç²ŸåºŠãæ±åæ§ãããã³å¹çã«ã€ãªããå¯èœæ§ããããŸãã
ãã€ããŒãã©ã¡ãŒã¿èª¿æŽã®èª²é¡
ãã€ããŒãã©ã¡ãŒã¿ã®æé©åã¯ãããã€ãã®èª²é¡ããããããç°¡åãªäœæ¥ã§ã¯ãããŸããã
- 髿¬¡å ã®æ¢çŽ¢ç©ºéïŒå¯èœãªãã€ããŒãã©ã¡ãŒã¿ã®çµã¿åããã®ç©ºéã¯åºå€§ã«ãªãå¯èœæ§ããããç¹ã«å€ãã®ãã€ããŒãã©ã¡ãŒã¿ãæã€ã¢ãã«ã®å Žåã«åœãŠã¯ãŸããŸãã
- éåžæé©åïŒãã€ããŒãã©ã¡ãŒã¿ãšã¢ãã«ã®ããã©ãŒãã³ã¹ã®é¢ä¿ã¯ãå€ãã®å Žåéåžã§ãããã°ããŒãã«ãªæé©å€ãèŠã€ããã®ãå°é£ã§ãã
- é«äŸ¡ãªè©äŸ¡ïŒãã€ããŒãã©ã¡ãŒã¿æ§æãè©äŸ¡ããã«ã¯ãã¢ãã«ã®ãã¬ãŒãã³ã°ãšæ€èšŒãå¿ èŠã§ãããç¹ã«è€éãªã¢ãã«ãå€§èŠæš¡ãªããŒã¿ã»ããã®å Žåãèšç®ã³ã¹ããé«ããªãå¯èœæ§ããããŸãã
- ãã€ãºã®å€ãè©äŸ¡ïŒã¢ãã«ã®ããã©ãŒãã³ã¹ã¯ãããŒã¿ãµã³ããªã³ã°ãåæåãªã©ã®ã©ã³ãã ãªèŠå ã®åœ±é¿ãåããå¯èœæ§ãããããã€ããŒãã©ã¡ãŒã¿æ§æã®ãã€ãºã®å€ãè©äŸ¡ã«ã€ãªãããŸãã
ã°ãªãããµãŒããã©ã³ãã ãµãŒããªã©ã®åŸæ¥ã®æ¹æ³ã¯ãç¹ã«é«æ¬¡å ã®æ¢çŽ¢ç©ºéãé«äŸ¡ãªè©äŸ¡ãæ±ãå Žåãéå¹ççã§æéããããããšããããããŸãã
ãã€ãºæé©åã®ç޹ä»
ãã€ãºæé©åã¯ã確çã¢ãã«ã«åºã¥ãæé©åææ³ã§ããã颿°ãéåžåããã€ãºãå€ããããã³è©äŸ¡ã«ã³ã¹ãããããå Žåã§ããç®ç颿°ã®ã°ããŒãã«ãªæé©å€ãå¹ççã«èŠã€ããããšãç®çãšããŠããŸãããã€ãºã®å®çãå©çšããŠã芳枬ãããããŒã¿ã§ç®ç颿°ã«é¢ããäºåç¥èãæŽæ°ããæé©ãªãã€ããŒãã©ã¡ãŒã¿æ§æã®æ¢çŽ¢ãã¬ã€ãããããã«äœ¿çšãããäºåŸååžãäœæããŸãã
äž»ãªæŠå¿µ
- ãµãã²ãŒãã¢ãã«ïŒç®ç颿°ãè¿äŒŒãã確çã¢ãã«ïŒéåžžã¯ã¬ãŠã¹éçšïŒãæ¢çŽ¢ç©ºéå ã®åãã€ã³ãã§å¯èœãªé¢æ°å€ã®ååžãæäŸãã颿°ã®åäœã«é¢ããäžç¢ºå®æ§ãå®éåããããšãã§ããŸãã
- ç²åŸé¢æ°ïŒè©äŸ¡ããæ¬¡ã®ãã€ããŒãã©ã¡ãŒã¿æ§æã®æ¢çŽ¢ãã¬ã€ããã颿°ãæ¢çŽ¢ïŒæ¢çŽ¢ç©ºéã®æªèžé åã§ã®æ¢çŽ¢ïŒãšæŽ»çšïŒæœåšæ§ã®é«ãé åãžã®çŠç¹ïŒã®ãã©ã³ã¹ãåããŸãã
- ãã€ãºã®å®çïŒèŠ³æž¬ãããããŒã¿ã§ãµãã²ãŒãã¢ãã«ãæŽæ°ããããã«äœ¿çšãããŸããç®ç颿°ã«é¢ããäºåç¥èãšããŒã¿ããã®å°€åºŠæ å ±ãçµã¿åãããŠãäºåŸååžãçæããŸãã
ãã€ãºæé©åããã»ã¹
ãã€ãºæé©åããã»ã¹ã¯ã次ã®ããã«èŠçŽã§ããŸãã- åæåïŒã©ã³ãã ã«éžæãããããã€ãã®ãã€ããŒãã©ã¡ãŒã¿æ§æã§ç®ç颿°ãè©äŸ¡ããŸãã
- ãµãã²ãŒãã¢ãã«ã®æ§ç¯ïŒèŠ³æž¬ãããããŒã¿ã«ãµãã²ãŒãã¢ãã«ïŒã¬ãŠã¹éçšãªã©ïŒãé©åãããŸãã
- ç²åŸé¢æ°ã®æé©åïŒãµãã²ãŒãã¢ãã«ã䜿çšããŠç²åŸé¢æ°ãæé©åããŸããããã«ãããè©äŸ¡ããæ¬¡ã®ãã€ããŒãã©ã¡ãŒã¿æ§æãææ¡ãããŸãã
- ç®ç颿°ã®è©äŸ¡ïŒææ¡ããããã€ããŒãã©ã¡ãŒã¿æ§æã§ç®ç颿°ãè©äŸ¡ããŸãã
- ãµãã²ãŒãã¢ãã«ã®æŽæ°ïŒæ°ãã芳枬å€ã§ãµãã²ãŒãã¢ãã«ãæŽæ°ããŸãã
- ç¹°ãè¿ãïŒåæ¢åºæºãæºãããããŸã§ïŒããšãã°ãæå€§ååŸ©åæ°ãç®æšããã©ãŒãã³ã¹ã®éæïŒãã¹ããã3ã5ãç¹°ãè¿ããŸãã
ã¬ãŠã¹éçšïŒGPïŒã®çè§£
ã¬ãŠã¹éçšã¯ã颿°ãã¢ãã«åããäžç¢ºå®æ§ãå®éåããããã®åŒ·åãªããŒã«ã§ãããããã¯ãæ¢çŽ¢ç©ºéå ã®åãã€ã³ãã§å¯èœãªé¢æ°å€ã®ååžãæäŸããæ©èœãããããããã€ãºæé©åã®ãµãã²ãŒãã¢ãã«ãšããŠãã䜿çšãããŸãã
ã¬ãŠã¹éçšã®äž»ãªç¹æ§
- 颿°å šäœã®ååžïŒã¬ãŠã¹éçšã¯ãå¯èœãªé¢æ°å šäœã®ç¢ºçååžãå®çŸ©ããŸãã
- å¹³åãšå ±åæ£ã«ãã£ãŠå®çŸ©ãããŸãïŒã¬ãŠã¹éçšã¯ããã®å¹³å颿°m(x)ãšå ±åæ£é¢æ°k(x, x')ã«ãã£ãŠå®å šã«æå®ãããŸããå¹³å颿°ã¯åãã€ã³ãã§ã®é¢æ°ã®æåŸ å€ã衚ããå ±åæ£é¢æ°ã¯ç°ãªããã€ã³ãã§ã®é¢æ°å€éã®çžé¢é¢ä¿ãèšè¿°ããŸãã
- ã«ãŒãã«é¢æ°ïŒã«ãŒãã«é¢æ°ãšãåŒã°ããå ±åæ£é¢æ°ã¯ãã¬ãŠã¹éçšãããµã³ããªã³ã°ããã颿°ã®æ»ããããšåœ¢ç¶ã決å®ããŸããäžè¬çãªã«ãŒãã«é¢æ°ã«ã¯ãååŸåºåºé¢æ°ïŒRBFïŒã«ãŒãã«ãMatérnã«ãŒãã«ãããã³ç·åœ¢ã«ãŒãã«ãå«ãŸããŸãã
- äºåŸæšè«ïŒèŠ³æž¬ãããããŒã¿ãäžãããããšããã€ãºã®å®çã䜿çšããŠã¬ãŠã¹éçšãæŽæ°ãã颿°å šäœã®äºåŸååžãååŸã§ããŸãããã®äºåŸååžã¯ãããŒã¿ã芳枬ããåŸã®é¢æ°ã®åäœã«é¢ããæŽæ°ãããä¿¡é ŒåºŠã衚ããŸãã
ãã€ãºæé©åã«ãããã¬ãŠã¹éçšã®äœ¿ç𿹿³
ãã€ãºæé©åã§ã¯ãã¬ãŠã¹éçšã¯ç®ç颿°ãã¢ãã«åããããã«äœ¿çšãããŸããGPã¯ãåãã€ããŒãã©ã¡ãŒã¿æ§æã§å¯èœãªé¢æ°å€ã®ååžãæäŸãã颿°ã®åäœã«é¢ããäžç¢ºå®æ§ãå®éåããããšãã§ããŸãããã®äžç¢ºå®æ§ã¯ãæé©ãªãã€ããŒãã©ã¡ãŒã¿æ§æã®æ¢çŽ¢ãã¬ã€ãããããã«ç²åŸé¢æ°ã«ãã£ãŠäœ¿çšãããŸãã
ããšãã°ããã¥ãŒã©ã«ãããã¯ãŒã¯ã®åŠç¿çã調æŽããŠãããšããŸããã¬ãŠã¹éçšã¯ãåŠç¿çãšãããã¯ãŒã¯ã®æ€èšŒç²ŸåºŠãšã®é¢ä¿ãã¢ãã«åããŸããåŠç¿çããšã«å¯èœãªæ€èšŒç²ŸåºŠã®ååžãæäŸããããŸããŸãªåŠç¿çã®å¯èœæ§ãè©äŸ¡ããæé©ãªå€ã®æ¢çŽ¢ãã¬ã€ãããããšãã§ããŸãã
ç²åŸé¢æ°ïŒæ¢çŽ¢ãšæŽ»çšã®ãã©ã³ã¹
ç²åŸé¢æ°ã¯ãè©äŸ¡ããæ¬¡ã®ãã€ããŒãã©ã¡ãŒã¿æ§æã®æ¢çŽ¢ãã¬ã€ãããããšã«ããããã€ãºæé©åã«ãããŠéèŠãªåœ¹å²ãæãããŸããæ¢çŽ¢ïŒæ¢çŽ¢ç©ºéã®æªèžé åã§ã®æ¢çŽ¢ïŒãšæŽ»çšïŒæœåšæ§ã®é«ãé åãžã®çŠç¹ïŒã®ãã©ã³ã¹ãåããŸãããã€ãºæé©åã§ã¯ãããã€ãã®ç²åŸé¢æ°ãäžè¬çã«äœ¿çšãããŸãã
- æ¹å確çïŒPIïŒïŒç¹å®ã®ãã€ããŒãã©ã¡ãŒã¿æ§æã§ã®ç®ç颿°å€ãããããŸã§ã«èŠ³æž¬ãããæé©ãªå€ãããåªããŠãã確çãPIã¯ãæœåšæ§ã®é«ãé åã«çŠç¹ãåœãŠãããšã«ãããæŽ»çšãåªå ããŸãã
- æåŸ æ¹åïŒEIïŒïŒç¹å®ã®ãã€ããŒãã©ã¡ãŒã¿æ§æã§ã®ç®ç颿°å€ãããããŸã§ã«èŠ³æž¬ãããæé©ãªå€ãããåªããŠãããšäºæ³ãããéãEIã¯ãPIãšæ¯èŒããŠãæ¢çŽ¢ãšæŽ»çšã®ãã©ã³ã¹ã®åããã¢ãããŒããæäŸããŸãã
- äžéä¿¡é ŒéçïŒUCBïŒïŒç®ç颿°ã®äºæž¬å¹³åãšããµãã²ãŒãã¢ãã«ã®äžç¢ºå®æ§ã«åºã¥ãäžéä¿¡é Œéçãçµã¿åãããç²åŸé¢æ°ãUCBã¯ãäžç¢ºå®æ§ã®é«ãé åãåªå ããããšã«ãããæ¢çŽ¢ãåªå ããŸãã
é©åãªç²åŸé¢æ°ã®éžæ
ç²åŸé¢æ°ã®éžæã¯ãç¹å®ã®åé¡ãšãæ¢çŽ¢ãšæŽ»çšã®éã®æãŸãããã©ã³ã¹ã«ãã£ãŠç°ãªããŸããç®ç颿°ãæ¯èŒçæ»ããã§é©åã«åäœããå ŽåãæŽ»çšãåªå ããç²åŸé¢æ°ïŒããšãã°ãPIïŒãé©ããŠããå ŽåããããŸãããã ããç®ç颿°ãéåžžã«éåžåãŸãã¯ãã€ãºãå€ãå Žåãæ¢çŽ¢ãåªå ããç²åŸé¢æ°ïŒããšãã°ãUCBïŒããã广çã§ããå¯èœæ§ããããŸãã
äŸïŒç»ååé¡ã®ããã®æ·±å±€åŠç¿ã¢ãã«ã®ãã€ããŒãã©ã¡ãŒã¿ãæé©åããŠãããšããŸããæé©ãªãã€ããŒãã©ã¡ãŒã¿æ§æã®åææšå®å€ãè¯å¥œã§ããå Žåã¯ãæåŸ æ¹åãªã©ã®ç²åŸé¢æ°ãéžæããŠã¢ãã«ã埮調æŽããå¯èœãªéãæé«ã®ããã©ãŒãã³ã¹ãå®çŸããããšãã§ããŸããäžæ¹ãæé©ãªæ§æãäžæãªå Žåã¯ãäžéä¿¡é Œéçãªã©ã®ç²åŸé¢æ°ãéžæããŠããã€ããŒãã©ã¡ãŒã¿ç©ºéã®ããŸããŸãªé åãæ¢çŽ¢ããæœåšçã«ããåªãããœãªã¥ãŒã·ã§ã³ãçºèŠããããšãã§ããŸãã
ãã€ãºæé©åã®å®è£
Pythonã§ãã€ãºæé©åãå®è£ ããããã«å©çšã§ããã©ã€ãã©ãªãšãã¬ãŒã ã¯ãŒã¯ãããã€ããããŸãã
- Scikit-optimize (skopt): å¹ åºããã€ãºæé©åã¢ã«ãŽãªãºã ãšç²åŸé¢æ°ãæäŸããäžè¬çãªPythonã©ã€ãã©ãªãScikit-learnããã®ä»ã®æ©æ¢°åŠç¿ã©ã€ãã©ãªãšäºææ§ããããŸãã
- GPyOpt: ã¬ãŠã¹éçšã¢ãã«ã«çŠç¹ãåœãŠãå€ç®çæé©åãå¶çŽä»ãæé©åãªã©ã®é«åºŠãªæ©èœãæäŸãããã€ãºæé©åã©ã€ãã©ãªã
- BayesianOptimization: ã·ã³ãã«ã§äœ¿ãããããã€ãºæé©åã©ã€ãã©ãªã§ãåå¿è ã«é©ããŠããŸãã
Scikit-optimize (skopt) ã䜿çšããäŸ
Scikit-optimizeã䜿çšããŠããµããŒããã¯ã¿ãŒãã·ã³ïŒSVMïŒåé¡åã®ãã€ããŒãã©ã¡ãŒã¿ãæé©åããæ¹æ³ã®äŸã次ã«ç€ºããŸãã
```python from skopt import BayesSearchCV from sklearn.svm import SVC from sklearn.datasets import load_iris from sklearn.model_selection import train_test_split # IrisããŒã¿ã»ãããããŒãããŸã iris = load_iris() X_train, X_test, y_train, y_test = train_test_split(iris.data, iris.target, test_size=0.2, random_state=42) # ãã€ããŒãã©ã¡ãŒã¿æ¢çŽ¢ç©ºéãå®çŸ©ããŸã param_space = { 'C': (1e-6, 1e+6, 'log-uniform'), 'gamma': (1e-6, 1e+1, 'log-uniform'), 'kernel': ['rbf'] } # ã¢ãã«ãå®çŸ©ããŸã model = SVC() # ãã€ãºæé©åæ¢çŽ¢ãå®çŸ©ããŸã opt = BayesSearchCV( model, param_space, n_iter=50, # ååŸ©åæ° cv=3 # 亀差æ€èšŒã®å岿° ) # æé©åãå®è¡ããŸã opt.fit(X_train, y_train) # æé©ãªãã©ã¡ãŒã¿ãšã¹ã³ã¢ãåºåããŸã print("Best parameters: %s" % opt.best_params_) print("Best score: %s" % opt.best_score_) # ãã¹ãã»ããã§ã¢ãã«ãè©äŸ¡ããŸã accuracy = opt.score(X_test, y_test) print("Test accuracy: %s" % accuracy) ```ãã®äŸã§ã¯ãScikit-optimizeã䜿çšããŠããã€ããŒãã©ã¡ãŒã¿æ¢çŽ¢ç©ºéãå®çŸ©ããã¢ãã«ãå®çŸ©ãããã€ãºæé©åæ¢çŽ¢ãå®è¡ããæ¹æ³ã瀺ããŸãã`BayesSearchCV`ã¯ã©ã¹ã¯ãã¬ãŠã¹éçšã¢ããªã³ã°ãšç²åŸé¢æ°æé©åãèªåçã«åŠçããŸãããã®ã³ãŒãã§ã¯ã`C`ããã³`gamma`ãã©ã¡ãŒã¿ã«ãã°äžæ§ååžã䜿çšããŠããŸããããã¯ãæ°æ¡ã«ããã£ãŠå€åããå¯èœæ§ã®ãããã©ã¡ãŒã¿ã«é©ããŠããããšããããããŸãã`n_iter`ãã©ã¡ãŒã¿ã¯ååŸ©åæ°ãå¶åŸ¡ããå®è¡ãããæ¢çŽ¢ã®éãæ±ºå®ããŸãã`cv`ãã©ã¡ãŒã¿ã¯ãåãã€ããŒãã©ã¡ãŒã¿æ§æãè©äŸ¡ããããã«äœ¿çšããã亀差æ€èšŒã®å岿°ãæå®ããŸãã
ãã€ãºæé©åã®é«åºŠãªãã¯ããã¯
ããã€ãã®é«åºŠãªãã¯ããã¯ã¯ããã€ãºæé©åã®ããã©ãŒãã³ã¹ãããã«åäžãããããšãã§ããŸãã
- å€ç®çæé©åïŒè€æ°ã®ç®çãåæã«æé©åããŸãïŒããšãã°ã粟床ãšãã¬ãŒãã³ã°æéïŒã
- å¶çŽä»ãæé©åïŒãã€ããŒãã©ã¡ãŒã¿ã«é¢ããå¶çŽïŒããšãã°ãäºç®å¶çŽãå®å šå¶çŽïŒã«åŸã£ãŠç®ç颿°ãæé©åããŸãã
- 䞊åãã€ãºæé©åïŒè€æ°ã®ãã€ããŒãã©ã¡ãŒã¿æ§æã䞊è¡ããŠè©äŸ¡ããŠãæé©åããã»ã¹ãé«éåããŸãã
- 転移åŠç¿ïŒä»¥åã®æé©åå®è¡ããã®ç¥èãæŽ»çšããŠãæ°ããåé¡ã®æé©åããã»ã¹ãå éããŸãã
- ãã³ãã£ããããŒã¹ã®æé©åïŒãã€ãºæé©åããã³ãã£ããã¢ã«ãŽãªãºã ãšçµã¿åãããŠããã€ããŒãã©ã¡ãŒã¿ç©ºéãå¹ççã«æ¢çŽ¢ããŸãã
äŸïŒäžŠåãã€ãºæé©å
䞊åãã€ãºæé©åã¯ãç¹ã«ãã€ããŒãã©ã¡ãŒã¿æ§æã®è©äŸ¡ã«èšç®ã³ã¹ãããããå Žåããã€ããŒãã©ã¡ãŒã¿èª¿æŽã«å¿ èŠãªæéãå€§å¹ ã«ççž®ã§ããŸããå€ãã®ã©ã€ãã©ãªã¯äžŠååã®ããã®çµã¿èŸŒã¿ãµããŒããæäŸããŠããããŸãã¯Pythonã®`concurrent.futures`ãªã©ã®ã©ã€ãã©ãªã䜿çšããŠæåã§å®è£ ããããšãã§ããŸãã
éèŠãªã¢ã€ãã¢ã¯ãç²åŸé¢æ°ã«ãã£ãŠææ¡ãããè€æ°ã®ãã€ããŒãã©ã¡ãŒã¿æ§æãåæã«è©äŸ¡ããããšã§ããããã«ã¯ã䞊åè©äŸ¡ãæé©åããã»ã¹ã«é©åã«çµã¿èŸŒãŸããããã«ããµãã²ãŒãã¢ãã«ãšç²åŸé¢æ°ãæ³šææ·±ã管çããå¿ èŠããããŸãã
äŸïŒå¶çŽä»ããã€ãºæé©å
å€ãã®çŸå®ã®ã·ããªãªã§ã¯ããã€ããŒãã©ã¡ãŒã¿èª¿æŽã¯å¶çŽãåããŸããããšãã°ãã¢ãã«ã®ãã¬ãŒãã³ã°ã«å©çšã§ããäºç®ãéãããŠããå Žåããã¢ãã«ãç¹å®ã®å®å šèŠä»¶ãæºãããŠããããšã確èªããå¿ èŠãããå ŽåããããŸãã
å¶çŽä»ããã€ãºæé©åãã¯ããã¯ã䜿çšããŠããããã®å¶çŽãæºãããªããç®ç颿°ãæé©åã§ããŸãããããã®ãã¯ããã¯ã«ã¯éåžžãå¶çŽãç²åŸé¢æ°ãŸãã¯ãµãã²ãŒãã¢ãã«ã«çµã¿èŸŒãããšãå«ãŸããŸãã
ãã€ãºæé©åã®å©ç¹ã𿬠ç¹
å©ç¹
- å¹çïŒãã€ãºæé©åã¯ãéåžžãã°ãªãããµãŒããã©ã³ãã ãµãŒããªã©ã®åŸæ¥ã®æ¹æ³ãšæ¯èŒããŠãç®ç颿°ã®è©äŸ¡åæ°ãå°ãªãæžã¿ãé«äŸ¡ãªé¢æ°ã®æé©åã«å¹ççã§ãã
- éåžæ§ãåŠçããŸãïŒãã€ãºæé©åã¯ãæ©æ¢°åŠç¿ã§äžè¬çãªéåžç®ç颿°ãåŠçã§ããŸãã
- äžç¢ºå®æ§ãå®éåããŸãïŒãã€ãºæé©åã¯ãç®ç颿°ã«é¢ããäžç¢ºå®æ§ã®å°ºåºŠãæäŸããŸããããã¯ãæé©åããã»ã¹ãçè§£ããæ å ±ã«åºã¥ããæææ±ºå®ãè¡ãã®ã«åœ¹ç«ã¡ãŸãã
- é©å¿æ§ïŒãã€ãºæé©åã¯ãç®ç颿°ã®åœ¢ç¶ã«é©å¿ããæ¢çŽ¢ç©ºéã®ææãªé åã«çŠç¹ãåœãŠãŸãã
æ¬ ç¹
- è€éãïŒãã€ãºæé©åã¯ãã°ãªãããµãŒããã©ã³ãã ãµãŒããªã©ã®åçŽãªæ¹æ³ãšæ¯èŒããŠãå®è£ ãšçè§£ãè€éã«ãªãå¯èœæ§ããããŸãã
- èšç®ã³ã¹ãïŒãµãã²ãŒãã¢ãã«ã®æ§ç¯ãšæŽæ°ã®èšç®ã³ã¹ãã¯ãç¹ã«é«æ¬¡å ã®æ¢çŽ¢ç©ºéã§ã¯å€§ãããªãå¯èœæ§ããããŸãã
- äºåååžã«å¯ŸããæåºŠïŒãµãã²ãŒãã¢ãã«ã®äºåååžã®éžæã¯ããã€ãºæé©åã®ããã©ãŒãã³ã¹ã«åœ±é¿ãäžããå¯èœæ§ããããŸãã
- ã¹ã±ãŒã©ããªãã£ïŒãã€ãºæé©åã¯ãéåžžã«é«æ¬¡å ã®æ¢çŽ¢ç©ºéã«ã¹ã±ãŒã«ããã®ãé£ããå ŽåããããŸãã
ãã€ãºæé©åã䜿çšããå Žå
ãã€ãºæé©åã¯ã次ã®ã·ããªãªã«ç¹ã«é©ããŠããŸãã
- é«äŸ¡ãªè©äŸ¡ïŒç®ç颿°ã®è©äŸ¡ã«èšç®ã³ã¹ãããããå ŽåïŒããšãã°ã深局åŠç¿ã¢ãã«ã®ãã¬ãŒãã³ã°ïŒã
- éåžç®ç颿°ïŒãã€ããŒãã©ã¡ãŒã¿ãšã¢ãã«ã®ããã©ãŒãã³ã¹ã®é¢ä¿ãéåžã®å Žåã
- éãããäºç®ïŒæéãŸãã¯ãªãœãŒã¹ã®å¶çŽã«ãããè©äŸ¡ã®æ°ãå¶éãããŠããå Žåã
- 髿¬¡å ã®æ¢çŽ¢ç©ºéïŒæ¢çŽ¢ç©ºéã髿¬¡å ã§ãããã°ãªãããµãŒããã©ã³ãã ãµãŒããªã©ã®åŸæ¥ã®æ¹æ³ãéå¹çãªå Žåã
ããšãã°ããã€ãºæé©åã¯ãç³ã¿èŸŒã¿ãã¥ãŒã©ã«ãããã¯ãŒã¯ïŒCNNïŒããªã«ã¬ã³ããã¥ãŒã©ã«ãããã¯ãŒã¯ïŒRNNïŒãªã©ã®æ·±å±€åŠç¿ã¢ãã«ã®ãã€ããŒãã©ã¡ãŒã¿ã調æŽããããã«ãã䜿çšãããŸãããããã®ã¢ãã«ã®ãã¬ãŒãã³ã°ã¯èšç®ã³ã¹ããé«ãããã€ããŒãã©ã¡ãŒã¿ç©ºéãåºå€§ã«ãªãå¯èœæ§ãããããã§ãã
åŸæ¥ã®ãã€ããŒãã©ã¡ãŒã¿èª¿æŽãè¶ ããŠïŒAutoML
ãã€ãºæé©åã¯ãå€ãã®èªåæ©æ¢°åŠç¿ïŒAutoMLïŒã·ã¹ãã ã®ã³ã¢ã³ã³ããŒãã³ãã§ããAutoMLã¯ãããŒã¿ååŠçãç¹åŸŽãšã³ãžãã¢ãªã³ã°ãã¢ãã«éžæããã€ããŒãã©ã¡ãŒã¿èª¿æŽãªã©ãæ©æ¢°åŠç¿ãã€ãã©ã€ã³å šäœãèªååããããšãç®çãšããŠããŸãããã€ãºæé©åãä»ã®ãã¯ããã¯ãšçµ±åããããšã«ãããAutoMLã·ã¹ãã ã¯ãå¹ åºãã¿ã¹ã¯ã®æ©æ¢°åŠç¿ã¢ãã«ãèªåçã«æ§ç¯ããã³æé©åã§ããŸãã
ããã€ãã®AutoMLãã¬ãŒã ã¯ãŒã¯ãå©çšå¯èœã§ãã
- Auto-sklearn: ã¢ãã«éžæããã€ããŒãã©ã¡ãŒã¿èª¿æŽãªã©ãæ©æ¢°åŠç¿ãã€ãã©ã€ã³å šäœãæé©åããããã«ãã€ãºæé©åã䜿çšããAutoMLãã¬ãŒã ã¯ãŒã¯ã
- TPOT: æé©ãªæ©æ¢°åŠç¿ãã€ãã©ã€ã³ãæ€åºããããã«éºäŒçããã°ã©ãã³ã°ã䜿çšããAutoMLãã¬ãŒã ã¯ãŒã¯ã
- H2O AutoML: æ©æ¢°åŠç¿ããã»ã¹ãèªååããããã®å¹ åºãã¢ã«ãŽãªãºã ãšæ©èœãæäŸããAutoMLãã©ãããã©ãŒã ã
ã°ããŒãã«ãªäŸãšèæ ®äºé
ãã€ãºæé©åã®ååãšãã¯ããã¯ã¯ãããŸããŸãªå°åãæ¥çã«æ®éçã«é©çšã§ããŸãããã ããã°ããŒãã«ãªã³ã³ããã¹ãã§ãã€ãºæé©åãé©çšããå Žåã¯ã次ã®èŠçŽ ãèæ ®ããããšãéèŠã§ãã
- ããŒã¿ã®å€æ§æ§ïŒã¢ãã«ã®ãã¬ãŒãã³ã°ãšæ€èšŒã«äœ¿çšãããããŒã¿ããã°ããŒãã«ãªéå£ã代衚ããŠããããšã確èªããŸããããã«ã¯ãããŸããŸãªå°åãæåããããŒã¿ãåéããå¿ èŠãããå ŽåããããŸãã
- æåçé æ ®ïŒæé©åããã»ã¹ã®çµæãè§£éããéã«ã¯ãæåçãªéãã«æ³šæããŠãã ãããããšãã°ãæé©ãªãã€ããŒãã©ã¡ãŒã¿æ§æã¯ãæåçãªã³ã³ããã¹ãã«ãã£ãŠç°ãªãå ŽåããããŸãã
- æ³èŠå¶ã®éµå®ïŒã¢ãã«ãããŸããŸãªå°åã§é©çšããããã¹ãŠã®èŠå¶ã«æºæ ããŠããããšã確èªããŸããããšãã°ãäžéšã®å°åã§ã¯ãããŒã¿ã®ãã©ã€ãã·ãŒãšã»ãã¥ãªãã£ã«é¢ãã峿 ŒãªèŠå¶ãããå ŽåããããŸãã
- ã³ã³ãã¥ãŒãã£ã³ã°ã€ã³ãã©ã¹ãã©ã¯ãã£ïŒã³ã³ãã¥ãŒãã£ã³ã°ãªãœãŒã¹ã®å¯çšæ§ã¯ãå°åã«ãã£ãŠç°ãªãå ŽåããããŸããã¯ã©ãŠãããŒã¹ã®ãã©ãããã©ãŒã ã䜿çšããŠããã€ãºæé©åã«ååãªèšç®èœåãžã®ã¢ã¯ã»ã¹ãæäŸããããšãæ€èšããŠãã ããã
äŸïŒã°ããŒãã«ãªäžæ£æ€åºã·ã¹ãã ãéçºããŠããäŒç€Ÿã¯ããã€ãºæé©åã䜿çšããŠæ©æ¢°åŠç¿ã¢ãã«ã®ãã€ããŒãã©ã¡ãŒã¿ã調æŽããå ŽåããããŸããã¢ãã«ãããŸããŸãªå°åã§é©åã«æ©èœããããã«ããã«ã¯ãããŸããŸãªåœãæåããããŒã¿ãåéããå¿ èŠããããŸãããŸããæ¯åºãã¿ãŒã³ãäžæ£è¡çºã«ãããæåçãªéããèæ ®ããå¿ èŠããããŸããããã«ãåå°åã®ããŒã¿ãã©ã€ãã·ãŒèŠå¶ãéµå®ããå¿ èŠããããŸãã
çµè«
ãã€ãºæé©åã¯ããã€ããŒãã©ã¡ãŒã¿èª¿æŽã®ããã®åŒ·åã§å¹ççãªãã¯ããã¯ã§ããå¹çæ§ãéåžæ§ãåŠçããèœåãããã³äžç¢ºå®æ§ã®å®éåãªã©ãã°ãªãããµãŒããã©ã³ãã ãµãŒããªã©ã®åŸæ¥ã®æ¹æ³ãããããã€ãã®å©ç¹ããããŸãããã€ãºæé©åã®ååãšãã¯ããã¯ãçè§£ããããšã§ãæ©æ¢°åŠç¿ã¢ãã«ã®ããã©ãŒãã³ã¹ãå€§å¹ ã«åäžãããå¹ åºãã¢ããªã±ãŒã·ã§ã³ã§ããè¯ãçµæãéæã§ããŸããããŸããŸãªã©ã€ãã©ãªãç²åŸé¢æ°ãããã³é«åºŠãªãã¯ããã¯ã詊ããŠãç¹å®ã®åé¡ã«æé©ãªã¢ãããŒããèŠã€ããŠãã ãããAutoMLãé²åãç¶ããã«ã€ããŠããã€ãºæé©åã¯æ©æ¢°åŠç¿ããã»ã¹ãèªååããããå¹ åºã察象è ãå©çšã§ããããã«ããããã«ãŸããŸãéèŠãªåœ¹å²ãæããã§ããããã¢ãã«ã®ã°ããŒãã«ãªåœ±é¿ãèæ ®ãã代衚çãªããŒã¿ãçµã¿èŸŒã¿ãæœåšçãªãã€ã¢ã¹ã«å¯ŸåŠããããšã«ããã倿§ãªéå£ã«ãããä¿¡é Œæ§ãšå ¬å¹³æ§ã確ä¿ããŠãã ããã