arXiv subjects

Manuel Baltieri

Publications and source records attributed to Manuel Baltieri.

At least 19 recordsLinked to original sources

arXiv2026

World models of environment, agent and joint agent-environment systems

World models are a central component of model-based reinforcement learning. They are usually discussed in terms of what variables they predict, such as observations, rewards, states, latent or information states. We argue that there is a prior distinction: which channel they model. We consider three cases: the environment channel $O_{:} \mid A_{:}$, the agent channel $A_{:} \mid O_{:}$, and the realised joint process $(A, O)_{:}$, equivalently viewed as a channel with no inputs. Using computational mechanics, we define canonical predictive models for these three cases as $\epsilon$-transducers or $\epsilon$-machines. Canonical environment models recover standard predictive state representations, while the other two give analogous notions of canonical models for the agent and the joint system. We then build canonical support-restricted environment and agent models induced by closed-loop coupling, whose predictive equivalences range over continuations supported by the realised interaction. The key structural result is that canonical support-restricted environment states factor through the canonical joint causal states, and their transition structure is induced directly from the joint model; the agent-side construction is dual. Finally, we give a POMDP/controller example in which the unrestricted environment model has infinitely many states while the canonical support-restricted model induced by the coupling is finite. The framework clarifies what different world models are models of, and how coupling and support restriction can change their canonical predictive structure and complexity.

cs.AI↗

arXiv2026

Compositional Behavioral Semantics for State Abstraction in Reinforcement Learning

State abstraction plays a key role in scaling reinforcement learning to complex but structured systems. In studying such systems, a wide range of behavioral structures have been studied in reinforcement learning, including value functions, invariants, bisimulation relations, and behavioral metrics. However, a general principle for determining what structures are provably preserved under state abstraction is still lacking. In this paper, we present a unified framework for defining and analyzing behavioral structures in reinforcement learning. Our framework provides a compositional way to specify behavioral semantics based on local, one-step descriptions of system dynamics. Using this framework, we establish results showing how behavioral structures can be safely transferred between abstract and concrete systems. We further show how to construct quantitative metrics from logical behavioral semantics with soundness guarantees. Together, these results provide a principled foundation for reasoning about behaviors under state abstraction in reinforcement learning and offer reusable definition and proof principles for a broad class of behavioral structures in reinforcement learning.

cs.LG↗

arXiv2026

Bayesian updates from coalgebraic determinisation

The powerset construction is the classical determinisation procedure for nondeterministic finite automata. In the coalgebraic setting, this construction has been generalised to structured coalgebras, which are coalgebras equipped with extra data. For stochastic Moore machines over the distribution monad, a type of structured coalgebra, the determinisation construction induces a semantics assigning to each finite input word a distribution on the current output. This semantics is appropriate when only the current output matters, but it is too coarse for settings in which intermediate observations must also be taken into account, as is typical for agents solving POMDPs in control theory and reinforcement learning. In these contexts, agents need to condition on all realised observations, not just the final one, so to better plan for the future. This has been addressed from a category theoretic perspective through a procedure called ``unifilarisation'', which (in our context) takes a stochastic Mealy machine and produces a machine whose states are priors over the original state space and whose transitions are given by Bayesian filtering. Here we show that unifilarisation is an instance of coalgebraic determinisation. We work with Mealy machines over monads equipped with extra structure generalising the notion of the support of a distribution. We show that in this setting, unifilarisation arises from the general determinisation procedure. We then compare the resulting final coalgebra semantics with the Moore-style one. Instead of assigning only a distribution on current outputs to each finite input word, it yields causal stochastic behaviours, that is, families mapping input words to distributions on output words compatible with the ``causality'' constraint that outputs cannot depend on future inputs.

cs.LO↗

arXiv2025

Prior preferences in active inference agents: soft, hard, and goal shaping

Active inference proposes expected free energy as an objective for planning and decision-making to adequately balance exploitative and explorative drives in learning agents. The exploitative drive, or what an agent wants to achieve, is formalised as the Kullback-Leibler divergence between a variational probability distribution, updated at each inference step, and a preference probability distribution that indicates what states or observations are more likely for the agent, hence determining the agent's goal in a certain environment. In the literature, the questions of how the preference distribution should be specified and of how a certain specification impacts inference and learning in an active inference agent have been given hardly any attention. In this work, we consider four possible ways of defining the preference distribution, either providing the agents with hard or soft goals and either involving or not goal shaping (i.e., intermediate goals). We compare the performances of four agents, each given one of the possible preference distributions, in a grid world navigation task. Our results show that goal shaping enables the best performance overall (i.e., it promotes exploitation) while sacrificing learning about the environment's transition dynamics (i.e., it hampers exploration).

cs.AI↗

arXiv2025

From monoliths to modules: Decomposing transducers for efficient world modelling

World models have been recently proposed as sandbox environments in which AI agents can be trained and evaluated before deployment. While realistic world models often have high computational demands, this can often be alleviated by exploiting the fact that real-world scenarios tend to involve subcomponents that interact in a modular manner. In this paper, we explore this idea by developing a framework for decomposing complex world models represented by transducers, a class of models generalising POMDPs. Whereas the composition of transducers is well understood, our results clarify how to invert this process by deriving sub-transducers operating on distinct input-output subspaces, enabling parallelizable and interpretable alternatives to monolithic world modelling that can support distributed inference. Overall, these results lay groundwork for bridging the computational efficiency required for real-world inference and the structural transparency demanded by AI safety.

cs.AI↗

arXiv2025

A coalgebraic perspective on predictive processing

Predictive processing and active inference posit that the brain is a system performing Bayesian inference on the environment. By virtue of this, a prominent interpretation of predictive processing states that the generative model (a POMDP) encoded by the brain synchronises with the generative process (another POMDP) representing the environment while trying to explain what hidden properties of the world generated its sensory input. In this view, the brain is thought to become a copy of the environment. This claim has however been disputed, stressing the fact that a structural copy, or isomorphism as it is at times invoked to be, is not an accurate description of this process since the environment is necessarily more complex than the brain, and what matters is not the capacity to exactly recapitulate the veridical causal structure of the world. In this work, we make parts of this counterargument formal by using ideas from the theory of coalgebras, an abstract mathematical framework for dynamical systems that brings together work from automata theory, concurrency theory, probabilistic processes and other fields. To do so, we cast generative model and process, in the form of POMDPs, as coalgebras, and use maps between them to describe a form of consistency that goes beyond mere structural similarity, giving the necessary mathematical background to describe how different processes can be seen as behaviourally, rather than structurally, equivalent, i.e. how they can be seen as emitting the same observations, and thus minimise prediction error, over time without strict assumptions about structural similarity. In particular, we will introduce three standard notions of equivalence from the literature on coalgebras, evaluating them in the context of predictive processing and identifying the one closest to claims made by proponents of this framework.

q-bio.NC↗

arXiv2025

Active inference for action-unaware agents

Active inference is a formal approach to study cognition based on the notion that adaptive agents can be seen as engaging in a process of approximate Bayesian inference, via the minimisation of variational and expected free energies. Minimising the former provides an account of perceptual processes and learning as evidence accumulation, while minimising the latter describes how agents select their actions over time. In this way, adaptive agents are able to maximise the likelihood of preferred observations or states, given a generative model of the environment. In the literature, however, different strategies have been proposed to describe how agents can plan their future actions. While they all share the notion that some kind of expected free energy offers an appropriate way to score policies, sequences of actions, in terms of their desirability, there are different ways to consider the contribution of past motor experience to the agent's future behaviour. In some approaches, agents are assumed to know their own actions, and use such knowledge to better plan for the future. In other approaches, agents are unaware of their actions, and must infer their motor behaviour from recent observations in order to plan for the future. This difference reflects a standard point of departure in two leading frameworks in motor control based on the presence, or not, of an efference copy signal representing knowledge about an agent's own actions. In this work we compare the performances of action-aware and action-unaware agents in two navigations tasks, showing how action-unaware agents can achieve performances comparable to action-aware ones while at a severe disadvantage.

cs.AI↗

arXiv2025

A "good regulator theorem" for embodied agents

In a classic paper, Conant and Ashby claimed that "every good regulator of a system must be a model of that system." Artificial Life has produced many examples of systems that perform tasks with apparently no model in sight; these suggest Conant and Ashby's theorem doesn't easily generalise beyond its restricted setup. Nevertheless, here we show that a similar intuition can be fleshed out in a different way: whenever an agent is able to perform a regulation task, it is possible for an observer to interpret it as having "beliefs" about its environment, which it "updates" in response to sensory input. This notion of belief updating provides a notion of model that is more sophisticated than Conant and Ashby's, as well as a theorem that is more broadly applicable. However, it necessitates a change in perspective, in that the observer plays an essential role in the theory: models are not a mere property of the system but are imposed on it from outside. Our theorem holds regardless of whether the system is regulating its environment in a classic control theory setup, or whether it's regulating its own internal state; the model is of its environment either way. The model might be trivial, however, and this is how the apparent counterexamples are resolved.

cs.AI↗

arXiv2025

AI in a vat: Fundamental limits of efficient world modelling for agent sandboxing and interpretability

Recent work proposes using world models to generate controlled virtual environments in which AI agents can be tested before deployment to ensure their reliability and safety. However, accurate world models often have high computational demands that can severely restrict the scope and depth of such assessments. Inspired by the classic `brain in a vat' thought experiment, here we investigate ways of simplifying world models that remain agnostic to the AI agent under evaluation. By following principles from computational mechanics, our approach reveals a fundamental trade-off in world model construction between efficiency and interpretability, demonstrating that no single world model can optimise all desirable characteristics. Building on this trade-off, we identify procedures to build world models that either minimise memory requirements, delineate the boundaries of what is learnable, or allow tracking causes of undesirable outcomes. In doing so, this work establishes fundamental limits in world modelling, leading to actionable guidelines that inform core design choices related to effective agent evaluation.

cs.AI↗

arXiv2025

A Bayesian Interpretation of the Internal Model Principle

The internal model principle, originally proposed in the theory of control of linear systems, nowadays represents a more general class of results in control theory and cybernetics. The central claim of these results is that, under suitable assumptions, if a system (a controller) can regulate against a class of external inputs (from the environment), it is because the system contains a model of the system causing these inputs, which can be used to generate signals counteracting them. Similar claims on the role of internal models appear also in cognitive science, especially in modern Bayesian treatments of cognitive agents, often suggesting that a system (a human subject, or some other agent) models its environment to adapt against disturbances and perform goal-directed behaviour. It is however unclear whether the Bayesian internal models discussed in cognitive science bear any formal relation to the internal models invoked in standard treatments of control theory. Here, we first review the internal model principle and present a precise formulation of it using concepts inspired by categorical systems theory. This leads to a formal definition of ``model'' generalising its use in the internal model principle. Although this notion of model is not a priori related to the notion of Bayesian reasoning, we show that it can be seen as a special case of possibilistic Bayesian filtering. This result is based on a recent line of work formalising, using Markov categories, a notion of ``interpretation'', describing when a system can be interpreted as performing Bayesian filtering on an outside world in a consistent way.

math.OC↗

arXiv2024

Disentangled Representations for Causal Cognition

Complex adaptive agents consistently achieve their goals by solving problems that seem to require an understanding of causal information, information pertaining to the causal relationships that exist among elements of combined agent-environment systems. Causal cognition studies and describes the main characteristics of causal learning and reasoning in human and non-human animals, offering a conceptual framework to discuss cognitive performances based on the level of apparent causal understanding of a task. Despite the use of formal intervention-based models of causality, including causal Bayesian networks, psychological and behavioural research on causal cognition does not yet offer a computational account that operationalises how agents acquire a causal understanding of the world. Machine and reinforcement learning research on causality, especially involving disentanglement as a candidate process to build causal representations, represent on the one hand a concrete attempt at designing causal artificial agents that can shed light on the inner workings of natural causal cognition. In this work, we connect these two areas of research to build a unifying framework for causal cognition that will offer a computational perspective on studies of animal cognition, and provide insights in the development of new algorithms for causal reinforcement learning in AI.

cs.AI↗

arXiv2022

Hybrid Life: Integrating Biological, Artificial, and Cognitive Systems

Artificial life is a research field studying what processes and properties define life, based on a multidisciplinary approach spanning the physical, natural and computational sciences. Artificial life aims to foster a comprehensive study of life beyond "life as we know it" and towards "life as it could be", with theoretical, synthetic and empirical models of the fundamental properties of living systems. While still a relatively young field, artificial life has flourished as an environment for researchers with different backgrounds, welcoming ideas and contributions from a wide range of subjects. Hybrid Life is an attempt to bring attention to some of the most recent developments within the artificial life community, rooted in more traditional artificial life studies but looking at new challenges emerging from interactions with other fields. In particular, Hybrid Life focuses on three complementary themes: 1) theories of systems and agents, 2) hybrid augmentation, with augmented architectures combining living and artificial systems, and 3) hybrid interactions among artificial and biological systems. After discussing some of the major sources of inspiration for these themes, we will focus on an overview of the works that appeared in Hybrid Life special sessions, hosted by the annual Artificial Life Conference between 2018 and 2022.

cs.AI↗

arXiv2021

Kalman filters as the steady-state solution of gradient descent on variational free energy

The Kalman filter is an algorithm for the estimation of hidden variables in dynamical systems under linear Gauss-Markov assumptions with widespread applications across different fields. Recently, its Bayesian interpretation has received a growing amount of attention especially in neuroscience, robotics and machine learning. In neuroscience, in particular, models of perception and control under the banners of predictive coding, optimal feedback control, active inference and more generally the so-called Bayesian brain hypothesis, have all heavily relied on ideas behind the Kalman filter. Active inference, an algorithmic theory based on the free energy principle, specifically builds on approximate Bayesian inference methods proposing a variational account of neural computation and behaviour in terms of gradients of variational free energy. Using this ambitious framework, several works have discussed different possible relations between free energy minimisation and standard Kalman filters. With a few exceptions, however, such relations point at a mere qualitative resemblance or are built on a set of very diverse comparisons based on purported differences between free energy minimisation and Kalman filtering. In this work, we present a straightforward derivation of Kalman filters consistent with active inference via a variational treatment of free energy minimisation in terms of gradient descent. The approach considered here offers a more direct link between models of neural dynamics as gradient descent and standard accounts of perception and decision making based on probabilistic inference, further bridging the gap between hypotheses about neural implementation and computational principles in brain and behavioural sciences.

q-bio.NC↗

arXiv2020

Predictions in the eye of the beholder: an active inference account of Watt governors

Active inference introduces a theory describing action-perception loops via the minimisation of variational (and expected) free energy or, under simplifying assumptions, (weighted) prediction error. Recently, active inference has been proposed as part of a new and unifying framework in the cognitive sciences: predictive processing. Predictive processing is often associated with traditional computational theories of the mind, strongly relying on internal representations presented in the form of generative models thought to explain different functions of living and cognitive systems. In this work, we introduce an active inference formulation of the Watt centrifugal governor, a system often portrayed as the canonical "anti-representational" metaphor for cognition. We identify a generative model of a steam engine for the governor, and derive a set of equations describing "perception" and "action" processes as a form of prediction error minimisation. In doing so, we firstly challenge the idea of generative models as explicit internal representations for cognitive systems, suggesting that such models serve only as implicit descriptions for an observer. Secondly, we consider current proposals of predictive processing as a theory of cognition, focusing on some of its potential shortcomings and in particular on the idea that virtually any system admits a description in terms of prediction error minimisation, suggesting that this theory may offer limited explanatory power for cognitive systems. Finally, as a silver lining we emphasise the instrumental role this framework can nonetheless play as a mathematical tool for modelling cognitive architectures interpreted in terms of Bayesian (active) inference.

q-bio.NC↗

arXiv2020

On Kalman-Bucy filters, linear quadratic control and active inference

Linear Quadratic Gaussian (LQG) control is a framework first introduced in control theory that provides an optimal solution to linear problems of regulation in the presence of uncertainty. This framework combines Kalman-Bucy filters for the estimation of hidden states with Linear Quadratic Regulators for the control of their dynamics. Nowadays, LQG is also a common paradigm in neuroscience, where it is used to characterise different approaches to sensorimotor control based on state estimators, forward and inverse models. According to this paradigm, perception can be seen as a process of Bayesian inference and action as a process of optimal control. Recently, active inference has been introduced as a process theory derived from a variational approximation of Bayesian inference problems that describes, among others, perception and action in terms of (variational and expected) free energy minimisation. Active inference relies on a mathematical formalism similar to LQG, but offers a rather different perspective on problems of sensorimotor control in biological systems based on a process of biased perception. In this note we compare the mathematical treatments of these two frameworks for linear systems, focusing on their respective assumptions and highlighting their commonalities and technical differences.

q-bio.NC↗

arXiv2020

A Bayesian perspective on classical control

The connections between optimal control and Bayesian inference have long been recognised, with the field of stochastic (optimal) control combining these frameworks for the solution of partially observable control problems. In particular, for the linear case with quadratic functions and Gaussian noise, stochastic control has shown remarkable results in different fields, including robotics, reinforcement learning and neuroscience, especially thanks to the established duality of estimation and control processes. Following this idea we recently introduced a formulation of PID control, one of the most popular methods from classical control, based on active inference, a theory with roots in variational Bayesian methods, and applications in the biological and neural sciences. In this work, we highlight the advantages of our previous formulation and introduce new and more general ways to tackle some existing problems in current controller design procedures. In particular, we consider 1) a gradient-based tuning rule for the parameters (or gains) of a PID controller, 2) an implementation of multiple degrees of freedom for independent responses to different types of signals (e.g., two-degree-of-freedom PID), and 3) a novel time-domain formalisation of the performance-robustness trade-off in terms of tunable constraints (i.e., priors in a Bayesian model) of a single cost functional, variational free energy.

math.OC↗

arXiv2019

Scaling active inference

In reinforcement learning (RL), agents often operate in partially observed and uncertain environments. Model-based RL suggests that this is best achieved by learning and exploiting a probabilistic model of the world. 'Active inference' is an emerging normative framework in cognitive and computational neuroscience that offers a unifying account of how biological agents achieve this. On this framework, inference, learning and action emerge from a single imperative to maximize the Bayesian evidence for a niched model of the world. However, implementations of this process have thus far been restricted to low-dimensional and idealized situations. Here, we present a working implementation of active inference that applies to high-dimensional tasks, with proof-of-principle results demonstrating efficient exploration and an order of magnitude increase in sample efficiency over strong model-free baselines. Our results demonstrate the feasibility of applying active inference at scale and highlight the operational homologies between active inference and current model-based approaches to RL.

cs.LG↗

arXiv2019

Generative models as parsimonious descriptions of sensorimotor loops

The Bayesian brain hypothesis, predictive processing and variational free energy minimisation are typically used to describe perceptual processes based on accurate generative models of the world. However, generative models need not be veridical representations of the environment. We suggest that they can (and should) be used to describe sensorimotor relationships relevant for behaviour rather than precise accounts of the world.

q-bio.NC↗