Probabilistic Modeling
Many systems that we build in an engineering setting or models we create for statistical analysis are expected to perform some sort of reasoning to aid us in the process of decision-making. Such systems usually operate by taking in data and using this data to reach a conclusion. Models equipped with control systems may even be able to plan and execute sequences of actions to attain a goal based on this reasoning.
More sophisticated systems may be able to evaluate counterfactuals and reason about what would happen if a certain scenario were true and use this information to plan the optimal sequence of actions they need to take to reach a desired goal. To learn more about this, please see the page on active inference.
There are many examples of reasoning systems, whether natural or artificial, in use in the world around us everyday. Consider, for example, the task of a doctor who has knowledge of the human body. Her task is to receive verbal data from a patient, a list of symptoms, and reason her way to a conclusion about what disease is most likely given this available data, or evidence. Statistical models, even those as simple as linear regression models, may also be viewed as performing a rudimentary form of reasoning. Consider a linear model which can predict the price of a house given the number of bedrooms, number of bathrooms, and square footage. Once this model is trained, it could be presented with evidence or data - the number of bedrooms, bathrooms, and square footage of a house the model has never been trained on - and then reason to a conclusion about the expected price of the house.
Genius agents are useful for a particular type of reasoning, one that involves reasoning, or decision-making, under conditions of uncertainty. Processes we wish to model in the real world can be messy and the data we have to work with may be ambiguous. The modeling systems that we develop to reason about the world must contend with this ambiguity, changing circumstances, incomplete knowledge about the process being modeled, and poorly defined goals. Genius agents use a modeling paradigm known as Bayesian inference to allow them to reason under these uncertain conditions or go a step further and take optimal actions under uncertainty using active inference. These procedures produce models that are adaptable, interpretable, and come equipped with tools to measure the level of confidence in the model's predictions.
Genius agents represent the world in terms of a probabilistic graphical model. There are two specific types of graphical models we will focus on in these tutorials. The first is a discrete Bayesian network. Bayesian networks are useful to introduce the concepts and themes that underlie the logic behind Genius agents and the predictions they make. However, the key way to interface with Genius agents is through a type of probabilistic graphical model known as a factor graph which is generally equivalent to a Bayesian network.