Ensembles, one of my favorite topics. Now an ensemble is a team of models working together. So, I want you to imagine you're putting together a quiz team. You've got to week six of this course, we know you're clever. So maybe you're thinking, you're going to go solo, bit like Sheldon in Big Bang Theory. But everybody makes mistakes sometimes, and this is the core idea behind ensembles. Is each of your machine learning models will make mistakes on some kinds of data. So, you go and find three of your best friends. You're all studying similar subjects. You've all got similar hobbies and interests. Now on the plus side, the team huddles are going to be really harmonious. And in fact this is a good description of the way random forrest works. It's the same algorithm, there's just a little bit of random variation between each model. But is this team going to be varied enough? Throughout this course you will have seen various machine learning algorithms and already you know each of them have different strengths and weaknesses. For best results, you want your team, your ensemble, to be as diverse as possible. Now, all the H2O algorithms, when they're doing classification, give you a probability, a confidence in their prediction. So one simple way to implement an ensemble is to sum those confidences from each model. I have an example on my blog, and the URL will appear somewhere, showing how to do this in H2O. If you're doing a regression, you could simply take the mean of each of the predictions from your models. Or you may want to take the median or check them and get rid of outliers and so on. A more sophisticated way is to treat how should I combine the outputs from my models, has a machine learning problem. This is called a stacked ensemble. So you're building one model on top of your other models. And H2O comes with its own implementation of stacked ensembles. We're going to see how to use it in the next video. So just to sum up, ensembles of a diverse set of models is a key technique in data science. It might indicate you 1 or 2% improvement. But if you were going to be making a set of models anyway on your data, it's an improvement you can get for free.