Glossary

Active inference

Statistical methodology that underlies action selection and decision-making in Genius agents. Based on the work of neuroscientist Karl Friston and his collaborators. Genius agents employ the discrete state-space formulation of active inference in which latent states and actions are inferred in a Bayesian fashion on the basis of observations (i.e. data) available to the agent. Can be used for planning, decision-making, and control of external processes in the real world.

See the active inference tutorial for more details.

Agent

A system which has the ability to represent its environment as a probabilistic model, infer the probability of events of interest in the environment, take actions to control or manipulate the environment toward end goals, and learn from the feedback of its actions.

Bayesian inference

A method in probabilistic inference of inferring the probability of a parent variable given some observations about a child variable. This type of "reverse inference" operates backwards along the dependency line.

See the probabilistic inference tutorial for more detail.

Bayesian network

A graphical representation of a probabilistic model in which the model variables are nodes and the edges indicate dependency relationships among the variables.

See the discrete Bayesian networks tutorial for more detail.

Data

Also referred to as evidence in Bayesian networks or observations in the case of active inference and POMDP models. Data refers to specific instances of a variable's state that can be observed in the real world and gathered through controlled experiments and recorded by humans or machines (sensors).

Factor graph

A specific representation of a probabilistic model used by Genius agents. In this representation, factors of the probabilistic model represent either conditional or marginal probability distributions whose product produces the joint distribution defining the model. In factor graphs, these factors become factor nodes with associated probabilities that relate the variable nodes connected to the factors.

See the factor graphs tutorial for more detail.

Latent (hidden, unobserved) variable

Variables in a probabilistic model for which data is not available. We often say that such variables are "unobserved". Latent variables are useful to add into models because they may represent variables we wish to measure but do not have specific data available to do so. If latent variables are connected to other observable variables in the model, for which there is data available, it may be possible to infer the probabilities of latent variables despite missing this data.

See the section on latent variables in the probabilistic inference tutorial for more detail.

Likelihood

A function which represents the probability of the data over a range of values for some unobserved variable of interest. In the case of active inference, the likelihood represents the probability of observations (data) over the different possible, unobserved states of the system being modeled. In active inference models the likelihood function is used to infer latent states given the observations via Bayesian inference.

See the active inference tutorial for more details.

Markov property

An assumption in POMDP models that the previous state is all that is needed to make inferences about future states.

See the POMDP tutorial for more detail.

Model

Specifically a probabilistic model. A simplified representation of some real world process captured in the language of random variables that represent outcomes of clearly defined probabilistic events in the world and the probabilistic dependencies between them. Probabilistic models can be used for probabilistic inference of which active inference and Bayesian inference are special cases.

See the tutorial on probabilistic modeling for more detail.

Observations

See Data.

Parameter learning (training)

The process of learning the values of parameters (probabilities) in a factor graph by using data.

Partially observable Markov decision process (POMDP)

A probabilistic model with specific assumptions:

• The system being modeled is partially observable in the sense that some variables (states) are unobserved.

• The agent has the ability to make decisions about courses of action and execute them to modify or affect the environment.

• The model assumes that the past state of the environment has all the information that agent needs to decide on a course of action (the Markov property).

See the parameter learning tutorial for more detail.

Probabilistic inference

The process of inferring the probability of variables of interest in a probabilistic model after receiving new data.

See the probabilistic inference tutorial for more detail.

Probabilistic model

A set of random variables, the dependencies between these variables, and the associated probabilities of outcomes. The probabilistic model is intended to represent some real world process of interest in the language of probability.

See the tutorial on probabilistic modeling for more detail.

States

A special kind of latent variable used in POMDP models associated with active inference. States correspond to some unobserved variable in the real world that changes or transitions over time.

States may also refer to the categories of a random variable.

See the active inference tutorial for more detail.

VFG

A VERSES factor graph (VFG) is the name for model files created with Genius. VFG models created in or imported to Genius are in JSON format.

See the JSON model format page for more information.