We introduce a minimal model of a thermodynamic agent capable of maintaining far-from-equilibrium states by actively harvesting and storing free energy from its environment. Inspired by minimal models of autonomy like Bittorio (Varela et al., 1991), our agent —labelled Thermina —gives shape to a theoretical framework for studying the interplay between thermodynamics and autonomy. By analytically studying the nonequilibrium steady state of the system, we distinguish between regions of ‘autonomous’ states —sustaining themselves out-of-equilibrium by harvesting free energy from the environment— and regions of ‘non-autonomous’ states —close to thermodynamic equilibrium and with very low chances of gathering free energy. Furthermore, we inspect the adaptive mechanisms that allow an agent to regulate its interaction with the environment to robustly maintain its nonequilibrium state. Studying in detail the behaviour of the system, we aim to provide insights into the broader question of how thermodynamic processes contribute to the emergence and maintenance of complex, adaptive behaviour in natural and artificial systems.

Integrated information and variational inference provide influential mathematical frameworks in neuroscience. Yet, the understanding of the connection between the two is limited. Here, we study a minimal model to show how variational inference displays large integrated information for highly correlated target distributions, in contrast with alternative inference approaches like maximum likelihood estimation.

The entropy production rate reflects the dissipation of free energy in a nonequilibrium state, and it is necessary for many biological functions. Nevertheless, trivial systems can display large entropy production, and it is yet an open challenge to characterize the out-of-equilibrium states of living systems and their operational meaning. We present a way to decompose the entropy production rate of a system, capturing how much of it is generated from higher-order interactions between its components. Our method combines recent information-geometric decompositions of the entropy production rate with a hierarchical decomposition of forces into k -body stochastic interactions.

Current success of Artificial Intelligence (particularly in the application of Deep Learning techniques) is bringing some of its methods closer to Artificial Life and re-opening old questions, social fears and envisioned applications. The concept of autonomy has long guided research and progress in Artificial Life. We explore how this concept can contribute to evaluate the autonomy of contemporary AI systems.

There has been a revival of the notion of habit in the embodied and situated cognitive sciences. A habit can be understood as ‘a self-sustaining pattern of sensorimotor coordination that is formed when the stability of a particular mode of sensorimotor engagement is dynamically coupled with the stability of the mechanisms generating it’ (Barandiaran, 2008, p. 281). This view has inspired models of biologically-inspired homeostatic agents capable of establishing their own habits (Di Paolo and Iizuka, 2008). Despite recent achievements in this field, there is little written about how social habits can be established from this modelling perspective. We hypothesize that, when the stability of internal behavioural mechanisms is coupled to the stability of a behaviour and other agents are present during this behaviour, a social interdependence of behaviour takes place: a social habit is established. We provide evidence for our hypothesis with an evolutionary robotics simulation model of homeostatic plasticity in a phototactic behaviour. Agents evolved to couple internal homeostasis to behavioural fitness display social interdependencies in their behaviour. The social habit of these agents was not interrupted when blindness to phototactic stimuli was introduced as long as social perception remained active. This did not happen when internal homeostasis was not coupled to the fitness of the agent. The results allow us to propose a possible conjecture about the character of social habits and to offer a potential theoretical framework to understand how habits develop from neurodynamics to the level of social interaction.

Affordances are directly perceived environmental possibilities for action. Born within ecological psychology, they have been proposed to be one of the main building blocks to explain cognition from and embodied and situated perspective. Despite the interest, a formal definition of affordances in information theory terms that would allow to exploit their full potential in models of cognitive systems is still missing. We explore the challenge of quantifying affordances by using information-theoretical measures. Specifically, we propose that empowerment (i.e., information quantifying how much influence and control an agent has over the environment it can perceive) can be used to formally capture information about the possibilities for action (the range of possible behaviors of the agent in a given environment), which in some cases can constitute affordances. We test this idea in a minimal model reproducing some aspects of a classical example of body-scaled affordances: an agent passing through an aperture. We use empowerment measures to characterize the affordance of passing through the aperture. We find out that empowerment measures yield a similar transition to the one found in experimental data in humans in the specialized literature on ecological psychology. The exercise points to some limitations for formalizing affordances and allows us to pose questions regarding how affordances can be differentiated from more generic possibilities for action.

The concept of agency is of fundamental importance for Cognitive Science. However, usual definitions of agency are loose and the work to capture and measure it using mathematical tools is still in its infancy. Recently, the framework of integrated information theory has been proposed to capture the causal boundaries of biological autonomous systems. Here, we test measures of integrated information theory in a minimal model to test its capacity to identify and delimit an autonomous agent interacting with an environment. Doing so, we reformulate some aspects of current definitions of agency using insights from integrated information in our models. Specifically, we propose a redefinition of how we capture the ability of an agent to modulate its interaction with the environment in terms of the control of the emergent causal structure of the agent-environment system. In this way, we propose an operational definition of agency based on the capacity of a system to modulate its causal boundary, extending and reducing it by functionally open and closing sensorimotor loops, and coupling the agent to different environmental processes. This allows us to formulate a tentative measure for our definition of agency and test it in minimal models of sensorimotor interaction, which we test in a minimal agent evolved to solve a simple task.

The concept of autonomy is fundamental for understanding biological organization and the evolutionary transitions of living systems. Understanding how a system constitutes itself as an individual, cohesive, self-organized entity is a fundamental challenge for the understanding of life. However, it is generally a difficult task to determine whether the system or its environment has generated the correlations that allow an observer to trace the boundary of a living system as a coherent unit. Inspired by the framework of integrated information theory, we propose a measure of the level of integration of a system as the response of a system to partitions that introduce perturbations in the interaction between subsystems, without assuming the existence of a stationary distribution. With the goal of characterizing transitions in integrated information in the thermodynamic limit, we apply this measure to kinetic Ising models of infinite size using mean field techniques. Our findings suggest that, in order to preserve the integration of causal influences of a system as it grows in size, a living entity must be poised near critical points maximizing its sensitivity to perturbations in the interaction between subsystems. Moreover, we observe how such a measure is able to delimit an agent and its environment, being able to characterize simple instances of agent-environment asymmetries in which the agent has the ability to modulate its coupling with the environment.

A popular hypothesis suggests that the nervous system of different organisms, from neural tissue to whole brains, may operate at or near a critical point. During the last decade, maximum entropy techniques have allowed to go beyond merely finding statistical signatures of criticality, to models directly inferred from data recorded in neural cultures, providing stronger evidence of criticality in neural activity. Nevertheless, these modeling techniques are restricted to neural cultures and have not been extended to neural tissue in living organisms. In this paper, we extend this line of research by analyzing signatures of criticality in a pairwise maximum entropy model inferred from neural recordings of C. elegans during freely-moving locomotion. From the analysis of the inferred models we find some signatures of criticality, as a divergence of the heat capacity of the system. Other indicators, such as Zipf’s distributions, were not found. However, inspecting a similar analysis based in a 2D lattice Ising model we suggest that this could be due to the restricted number of samples in our data set. The availability of larger recordings of the C. elegans neural system during free locomotion could provide more conclusive results.

There exists extensive literature in cognitive science and psychology claiming the presence of indicators of criticality in several cognitive phenomena. However, the absence of quantitative models to understand how criticality produces the observed behavior makes it impossible to derive testable predictions. In particular, in active perception, the characterization of visual processes is frequently based only on the appearance of patterns in the analysis of experimental data without exploring their causes. Assuming a more formal viewpoint, we propose that statistical mechanics would be a general framework to connect active perception with a complex systems perspective. We show that this approach provides methods and experimental tools to measure the "thermodynamic properties" of perceptual processes in a simple visual task by identifying when a system is operating in a critical regime. Our model characterizes different perceptive modes to solve a visual illusion through thermodynamic regimes. Finally, we connect them with different perceptual strategies of exploiting directional symmetries.

This paper outlines a methodological approach for designing adaptive agents driving themselves near points of criticality. Using a synthetic approach we construct a conceptual model that, instead of specifying mechanistic requirements to generate criticality, exploits the maintenance of an organizational structure capable of reproducing critical behavior. Our approach exploits the well-known principle of universality, which classifies critical phenomena inside a few universality classes of systems independently of their specific mechanisms or topologies. In particular, we implement an artificial embodied agent controlled by a neural network maintaining a correlation structure randomly sampled from a lattice Ising model at a critical point. We evaluate the agent in two classical reinforcement learning scenarios: the Mountain Car benchmark and the Acrobot double pendulum, finding that in both cases the neural controller reaches a point of criticality, which coincides with a transition point between two regimes of the agent’s behaviour, maximizing the mutual information between neurons and sensorimotor patterns. Finally, we discuss the possible applications of this synthetic approach to the comprehension of deeper principles connected to the pervasive presence of criticality in biological and cognitive systems.