Supervised learning would be applied to: i) predicting children's height as a function of their age and ii) predicting the gender of a new manuscript’s author. In both cases, the problems are characterized by the presence of labeled data. The first scenario involves learning a predictive function based on a relationship between input features (age) and the continuous output (height). The second scenario involves training a model using pre-labeled data to predict categorical outcomes, i.e., classifying authors based on previous learning from gender-labeled articles.

Predicting the number of Microsoft shares to be traded tomorrow is characterized as a regression problem because the outcome is a continuous variable. In regression problems, the model forecasts a real value as opposed to assigning inputs to discrete categories. The objective here is to predict a specific quantity of shares which requires modeling of the relationship between historic trading volumes and future possible outcomes using continuous data points.

Machine learning can be classified into three main types: supervised learning, unsupervised learning, and reinforcement learning. Supervised learning utilizes labeled data to train algorithms, which means that the algorithm learns by example. It aims to predict outcomes for unseen data based on the models formed during the training phase. Unsupervised learning, on the other hand, deals with unlabeled data, allowing the algorithm to discover hidden patterns or intrinsic structures in the input data. Reinforcement learning involves learning what actions to take, given certain situations, to maximize a numerical reward signal. It is particularly useful in allowing machines to learn behavior in environments where they must carry out a sequence of operations to achieve a goal.

Reinforcement learning could outperform supervised or unsupervised learning in environments where sequential decision-making is essential, and outcomes of actions become apparent only through time. It is advantageous in problems like game playing, autonomous vehicle navigation, or robotic control, where the agent learns optimal actions through the maximization of reward signals over multiple trials. Unlike supervised or unsupervised approaches, reinforcement learning helps to understand the environment through trial and error, accommodating complex decision processes with long-term dependencies and non-stationary data.

Predicting whether a company will declare bankruptcy is considered a classification problem because the outcome variable is categorical. The problem requires the model to assign the input data to one of the predefined classes, in this case, 'bankruptcy' or 'no bankruptcy.' This binary classification helps categorize the companies based on the learning from past data of companies that faced similar financial conditions. It is different from regression problems where the output variable is continuous.

The categorization of a machine learning problem as classification or regression depends on the nature of the outcome variable. If the outcome is discrete and falls into predefined categories, the problem is a classification problem. If the outcome is a continuous and real-valued number, the problem falls under regression. Additional considerations involve the goal of the predictive model, whether to categorize or predict a precise value, and the data structure and labels available.

Unsupervised learning is useful in discovering sub-types of spam mail because it can analyze the patterns and similarities within the dataset without prior labeling. This technique can cluster examples using methods such as k-means or hierarchical clustering to identify hidden groupings or novel spam categories. These insights allow for building robust filtering systems that adapt to emerging spam patterns, thereby enhancing email security protocols effectively.

Linear regression might be used to predict a student's future academic performance based on previous grades because it models the relationship between a dependent variable and one or more independent variables. By examining how the number of 'A' grades in the freshman year influences those in the sophomore year, linear regression can provide an estimate for future grades. This approach assumes a linear relationship between the data points which can yield an predictive model useful in academic performance analysis.

Evaluating a model's performance in supervised learning focuses on accuracy, precision, recall, F1-score, and confusion matrices, where predictions are matched against true labels of a labeled dataset to determine correctness. In contrast, unsupervised learning evaluation involves metrics like cohesion and separation (silhouette score), within-cluster variance, and inter-cluster distance since labeled data is not available. Such unsupervised evaluation focuses on the quality and relevance of discovered patterns or groupings rather than explicit prediction accuracy.

The procedural differences between supervised and unsupervised learning algorithms primarily lie in the nature of the data used and the model training approach. Supervised learning utilizes labeled datasets, where each training example is paired with an output label, allowing the model to learn an input-output mapping. Model evaluation hinges on measured accuracy against known labels. Conversely, unsupervised learning works with unlabeled datasets and relies on patterns or groupings within the data itself, requiring the use of evaluation methods like silhouette scores that assess clustering quality. The choice of algorithm depends on the data structure and the desired output form — whether predictive or descriptive.