seminar notes

two improvements of Auto-Sklearn for AutoML approach:

1. included a meta-learning step
   • for finding good instantiations of machine learning frameworks
     从给出的ml frameworks中找到更可能在新数据集上表现的更好的framework实例
   • 具体的做法是：
     collect both performance data and a set of meta-features, i.e. characteristics of the dataset that can be computed efficiently and that help to determine which algorithm to use on a new dataset.
   • 影响：could warmstart the Bayesian optimization procedure
     • results in a considerable boost in efficiency
2. included an automated ensemble construction step
   one shortage of the previous Auto-ML procedure is that once the optimal algorithm (model) has been founded, all other models are discarded even if some of them perform almost as well as the optimal model, which is very wasteful and not very robust. That's why Auto-sklearn proposes a post-processing method that could store these models to construct an ensemble of them rather than just discarding them. This approach can improve any Bayesian hyperparameter optimization method.