Active Inference and Human--Computer Interaction

TL;DR

The paper addresses the need for a rigorous theory of interaction design capable of handling uncertainty and diversity in human users and contexts. It proposes Active Inference as a closed-loop, Bayesian framework where humans and computers are modeled as probabilistic agents with forward and observation models, trading off expected outcomes to minimise surprise. It outlines how AIF can be applied offline for design and offline analysis, online for real-time interaction, and in reflective configurations with mutual models, while detailing core constructs such as Markov blankets, forward models, and preference priors. The work highlights the predictive, explanatory, evaluative, and design-guiding potential of AIF for HCI, while candidly discussing computational, methodological, and software challenges and outlining practical next steps toward implementing AIF-based interactive systems that adapt to users and contexts in real time.

Abstract

Active Inference is a closed-loop computational theoretical basis for understanding behaviour, based on agents with internal probabilistic generative models that encode their beliefs about how hidden states in their environment cause their sensations. We review Active Inference and how it could be applied to model the human-computer interaction loop. Active Inference provides a coherent framework for managing generative models of humans, their environments, sensors and interface components. It informs off-line design and supports real-time, online adaptation. It provides model-based explanations for behaviours observed in HCI, and new tools to measure important concepts such as agency and engagement. We discuss how Active Inference offers a new basis for a theory of interaction in HCI, tools for design of modern, complex sensor-based systems, and integration of artificial intelligence technologies, enabling it to cope with diversity in human users and contexts. We discuss the practical challenges in implementing such Active Inference-based systems.

Figures (13)

• Figure 1: The overall structure of the paper.
• Figure 2: A schematic diagram of the Active Inference algorithm. An agents plans via a rollout process, scoring the tree of possible future states according to their information gain ("informativeness") and pragmatic value ("goodness") (See §\ref{['ref:min_surprise']} and App.\ref{['sec:tutorial']} for details). This is summarised into a single quantity for each imminent action, the Expected Free Energy (EFE) and an action is sampled to minimise this value. The agent then updates its belief states using a Bayesian update given the sampled action and its perceptions from sensing.
• Figure 3: The Active Inference representation of the entire human--computer interaction loop as a dyad of mutually interacting agents. Each agent can only perceive or act on the environment via those variables that form its Markov blanket. Each agent is embedded in the environment of the other, and its actions impinge upon its partner's Markov blanket only via this environment.
• Figure 4: Different application modes of Active Inference. Active Inference can be used to simulate user behaviour, or joint user-system behaviour ("offline"). Alternatively, systems can be built that apply AIF in the interaction loop ("online"). Increasing levels of sophistication involve additional nested active inference agents, as in the reflective agent that operates using an internal AIF simulator of a user, as shown in the lower box. E indicates environment. Dotted circles are non-AIF units.
• Figure 5: The environment in an AIF--HCI loop can be a) Something which is jointly controlled by a human and AIF system. b) A transmission medium c) Something which the user observes or controls via the AIF system