Abstract
The free energy principle and its corollary, active inference, were introduced by Karl Friston as an explanation embodied perception and action in neuroscience, and since, it has been used to address many other issues in different fields mainly related to cognitive science like learning, optimal decision, or interpersonal inference. Negotiation is a process where each negotiator has conflicting motivation is aiming to maximize his utility and where agreement is reached when the opposing interests are balanced. The purpose of this chapter is to illustrate how the free energy principle might be used through active inference in modeling a negotiation process based on an example of real life. The work is an attempt to bring together a dynamic logic framework with appropriate operators to consider motivation among agents on one hand and the active inference framework on the other hand.
TopBackground
In this chapter, active inference is considered in the context of cognitive agents and cognitivism. A cognitive agent is an agent that not only can learn about itself and about the environment it is interacting with but also imagine how the world would look like under different courses of actions and decide which action should be taken. Among these agents, rational agents are those which have preferences for advantageous outcomes. Cognitivism is a theoretical framework for understanding the way the mind perceives, acquires, remembers, processes, stores information and solves problems (Stavredes, 2011). Cognitivism is very present in artificial intelligence and, in particular, in learning theories and machine learning. Active inference explains how the agent decides which action to take by considering the course of actions that fulfills its prior beliefs about preferred observations. At the heart of this process, generative models play a central role. A generative model is an internal probabilistic representation of the agent’s beliefs on how hidden states of the environment relate to observations and how they transit to each other. Based on the generative model and the target priors, purposeful behavior emerges from variational free energy minimization which ensures that the agent avoids surprising states. As a matter of fact, active inference is a self-organizing process of action policy selection explained in terms of minimizing the free energy expected under a course of action. This is because under some ergodic assumptions, the long-term average of surprise is entropy and for a system or an agent to survive they must actively attempt to minimize their own entropy (Friston, Kilner, & Harrison, 2006; Millidge, 2019).
Key Terms in this Chapter
Free Energy: An information theoretic quantity that upper bounds the surprisal of some data with respect to some generative model.
Surprisal: A measure for the unexpectedness of an observation. It is expressed as the negative log-probability of the observation.
Cognitive Agent: An agent that executes a decision cycle in which it processes events and selects actions based on cognitive notions such as beliefs and goals.
Likelihood: The probability of observing data given the causes of those data.
Rational Agent: An agent that, based on a realistic model, has preferences for advantageous outcomes, and will seek to achieve them in a learning scenario.
Recognition Model: An approximate probability distribution of the causes of data as a result from inverting the generative model.
Ergodic: A process is ergodic if its long-term time-average converges to its ensemble average.
Prior: The probability distribution of the causes of data encoding beliefs about the causes before observing the data.
Generative Model: A probabilistic mapping from causes to observed data.