Evaluating performance
Measuring how good an algorithm is at doing its job is not always an easy task. Take the case of unsupervised learning: we expect a good unsupervised algorithm to unveil the most interesting and useful structures from data. The assessment on what makes them interesting or useful, however, will depend on your specific end goal and often requires some human judgment as well. In reinforcement learning, a good algorithm will be able to come back with a sizeable total reward, unlocking the opportunity to keep maximizing the return of our continuous interaction with the environment. Also in this case, the concept of reward will depend on a specific definition of value, determined by the case we are solving.
If we stay, instead, in the area of supervised learning, the performance evaluation is more straightforward: since our objective is to predict something (numbers or categories), we can assess the performance by measuring the differences between predicted and...