The Design and Composition of Structural Causal Decision Processes

Sebastian Benthall, Alan Lujan · 2026

Read The Design and Composition of Structural Causal Decision Processes by Sebastian Benthall, Alan Lujan (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-02681.txt.

Introduction

Sebastian Benthall and Alan Lujan define two new classes of causal models for decision-making agents, motivated by the economics of computing systems. Such systems have subsystems, and agents inside them face built-in limits on cognitive resources and discount future value. Standard causal influence models state links between variables and payoffs, but leave some of this structure out.

Their structural causal decision models add constrained decisions, where an agent's options can be limited by earlier causes, and open root variables with no assigned distribution. The models compose well, a property the authors prove useful. A decision process is then a repeating version with a discount variable. It is strictly more expressive than the standard partially observable decision model, because it does not assume rational belief formation and can model memory formation itself.

Catalog record

CREATORSebastian Benthall, Alan Lujan
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE77 KB
LICENSEhttp://creativecommons.org/licenses/by-sa/4.0/
RIGHTSCC BY SA
SOURCEarXiv

Source: https://arxiv.org/abs/2605.02681

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

The Design and Composition of Structural Causal Decision Processes Sebastian Benthall, Alan Lujan Abstract . We present two new classes of causal models of decision-making agents. Our approach is motivated by the needs of modeling the economics of computing systems. These systems are composed of subsystems and can exhibit endogenous limits on cognitive resources and value discounting. Structural Causal Decision Models (SCDMs) expand on Structural Causal Influence Models. Like SCIMs, they explicitly represent the causal relationships between model variables and the payoffs of agent decisions. Additionally, agent decisions can be constrained by their causal antecedents, and SCDMs can have open root variables for which no probability distribution or structural equation is given. We show that SCDMs have a well-defined and computationally useful property of composability. Building on SCDMs, we then define a Structural Causal Decision Process (SCDP) as a recurring SCDM with a discount variable. SCDPs benefit from the useful composition properties of SCDMs. Moreover, SCDPs are strictly more expressive than POMDPs because they do not assume rational belief formation. Indeed, an SCDP can endogenously model the memory-formation process, and is thus useful for modeling resource rational agents in dynamic settings. SCDPs are also capable of modeling variable discounting, a tool used widely in social scientific modeling. We pose that SCDPs are a useful framework for policy simulation for the digitial economy, mechanism design for information systems, and digital twin modeling of cyberinfrastructure. Computing systems today are composed of many subsystems, which can include technical components as well as natural and artificially intelligent agents. Interactions between these components and agents consist of flows of data, which is then stored and used intelligently by other actors. Realistic models of these systems must be decomposable and reflect endogenous limits on cognitive resources. This paper presents a formal framework for constructing such models. We position this as a new method for flexibly representing potentially intelligent agents in an economy ( 5 ) . Our approach is also consistent with structural causal modeling ( 43 ) approaches from computer science and statistics. We will also perform computational complexity analysis to demonstrate the value of composability to the performance of solution algorithms. We anticipate applications of this framework to include: • Policy simulation. Exploring the consequences of public policy on the digital economy. • System design. Modeling data flows in dynamic systems that include intelligent agents, with parameters that can be controlled for mechanism design. • Digital twin modeling. Realistic modeling that combines strategic actors with physics-informed neural networks (PINN) for scenario-based risk analysis. The framework developed in this paper stands on much prior work. Most closely, it builds to something similar to a dynamic influence model ( 38 ) . However, we build on more recent work on structural causal games ( 22 ) as well as the modeling concerns of computational economics, and more recent advances in deep learning, to arrive at a new design. In this paper, we will focus on the definition of composable, causal environments for a single agent. We see this as the first step. Future work will expand the definition to multi-agent environments, and address algorithms for learning the equilibrium strategies of the agents and using the models for statistical analysis. 0.1. Contributions Section 1 situates this article in prior literature and recent developments in computational economics and social science, AI safety, composable world models, and variable discounting. Section 2 introduces the Structural Causal Decision Model, a variation on Structural Causal Influence Models (see Definition 4 , ( 22 ) ) that supports constraints on decision variables and root variables with no probability distribution. We will show how SCDMs can be composed and decomposed, and how, in a special case of sequential decomposition, this readily improves the efficiency of solving for optimal decision rules. This is illustrated with 2-period consumption saving problems, one of which introduces habit-formation. Section 3 expands on the SCDM to show how it can be converted into a dynamic stochastic optimization problem. The new construct, a Structural Causal Decision Process (SCDP), can also be decomposed to improve solution efficiency. This is illustrated with a dynamic model that includes both consumption and portfolio allocation. Section 4 discusses the expressive potential of SCDPs, showing that they are more expressive than both MDPs and POMDPs. We demonstrate that in exceeding the expressiveness of POMDPs, SCDPs are able to model resource rationality by endogenizing cognitive constraints and tradeoffs of agents. Section 5 presents an adjustment to SCDPs that enables them to represent varying discount factors. We show that an SCDP can express models from macroeconomics where the discount factor, or patience, of consumers varies as a Markov process. In sum, this article builds up a new general formalism for representing dynamic optimization problems that is expressed in terms of Pearlian causation, has useful properties of compositionality, and can model resource rationality and variable or abstract discounting. We intend SCDMs and SCDPs to be a new computational class of models that is well-suited to social scientific modeling of complex sociotechnical systems, such as human-AI hybrid ecosystems. We present the formal framework and motivation for these models in anticipation of ongoing work that expands them to represent multi-agent systems, and develops efficient algorithms for fitting these models to data and solving them for optimal decision rules. The implementing project is called scikit-agent . It’s documentation 1 1 1 https://scikit-agent.github.io/scikit-agent/ and source code 2 2 2 https://github.com/scikit-agent/scikit-agent are available online. level agents – SCDP – – SCDM – SCM SCIM SCG CBN CID CG BN ID MAID BBayesian CCausal DDecision/Diagram GGame I M Influence MModel MAMulti-Agent NNetwork PProcess SStructural Figure 1. We extend the taxonomy of causal models from 22 , and in this paper introduce SCDM and SCDP. In the original diagram, the causal hierarchy (associational, interventional, and counterfactual) forms the vertical axis and the number of agents (0, 1, and N ) forms the horizontal axis. We add two additional elements to the causal hierarchy: functional, and dynamic. This paper introduces two new classes of causal models, Structural Causal Decision Models (SCDM) and Structural Causal Decision Processes (SCDP) that illustrate these steps in the hierarchy. 1. Prior work This work aims to synthesize recent developments in several areas of social and computer science as a foundation for asking and answering new kinds of questions about computing systems. 1.1. Economics and computational social science Agent-based modeling (ABM) has intrigued social scientists for decades. Computational tools and ideas from physics informed computational sociology and enabled early work on studying “artificial societies” ( 16 ) . ABM has since been understood to be a bridge between disciplines ( 4 ) . On the other hand, economics has typically viewed ABM with some skepticism, preferring models that are more explicitly decision-theoretic ( 8 ) . However, in their recent reformulation of the relationship between ABM and economics, 5 paint an inspiring picture of its current and future role within the field. The availability of ground truth micro-data and cheaper compute has made mathematical assumptions about aggregate or emergent level phenomena, which has been the mainstay of many earlier methods, less appealing than direct simulation of large populations of simulated agents. These simulations allow the modeler to relax strict and often unrealistic assumptions of

Read the complete text