Modèle Graphique Probabiliste
A Modèle Graphique Probabiliste (PGM) is a powerful framework used to describe the dependencies among random variables in a structured way using graphs. These models combine principles from théorie des probabilités and théorie des graphes, allowing for efficient representation and computation of complex joint probability distributions.
In PGMs, nodes typically represent random variables, and edges signify the probabilistic dependencies between them. There are two main types of PGMs: Dirigé Modèles Graphiques, also known as Bayesian Networks, and Modèles Graphiques Non Dirigés, commonly referred to as Markov Random Fields.
Dans un Réseau bayésien, the directed edges indicate a causal relationship, where the state of one variable can influence another. For example, in a medical diagnosis context, symptoms can be represented as nodes that are influenced by various diseases. This allows for reasoning about the likelihood of diseases given observed symptoms.
On the other hand, Markov Random Fields use undirected edges to represent symmetric relationships, where the distribution conjointe of the variables can be described by local interactions. These models are particularly useful in scenarios where the relationships are not strictly hierarchical, such as in image processing or social networks.
One of the key advantages of PGMs is their ability to efficiently handle uncertainty and make inferences about the variables. Techniques such as belief propagation and variational inference are commonly used for reasoning within these models. PGMs have applications in various domains, including machine learning, computer vision, traitement du langage naturel, and bioinformatics, making them a crucial tool for understanding and modeling complex systems.