New causal decision frameworks let economists model bounded rationality and variable discounting inside composable agent models; SCDPs are provably more expressive than POMDPs and suit policy and mechanism simulations for digital infrastructures.

The Design and Composition of Structural Causal Decision Processes

Sebastian Benthall, Alan Lujan · May 04, 2026

arxiv theoretical n/a evidence 8/10 relevance Source PDF

The paper introduces Structural Causal Decision Models and Structural Causal Decision Processes—composable causal frameworks that represent endogenous cognitive constraints, memory formation, and variable discounting—and shows they are strictly more expressive than POMDPs for modeling boundedly rational agents.

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 digital economy, mechanism design for information systems, and digital twin modeling of cyberinfrastructure.

Summary

Main Finding

The paper introduces two formal causal frameworks for decision-making agents—Structural Causal Decision Models (SCDMs) and Structural Causal Decision Processes (SCDPs)—that (a) extend Pearl-style structural causal and influence models to support constrained decisions and open root variables, (b) admit useful compositional properties that improve computational tractability, and (c) are strictly more expressive than MDPs/POMDPs (notably able to endogenize memory and cognitive constraints and represent variable discounting). The authors argue these models are well suited to policy simulation, mechanism design, and digital-twin modeling in socio-technical and AI-influenced economic systems.

Key Points

Definitions and structure
- SCDM: a DAG over endogenous and exogenous variables with endogenous variables partitioned into state (X), decision (D), and utility (U) nodes; a probability over exogenous variables; deterministic structural functions f for non-decision endogenous nodes; explicit constraints Γ on decisions; and support for root variables that may lack assigned distributions.
- SCDP: a dynamic (recurring) SCDM that includes an explicit discount variable; serves as the dynamic optimization/process-level extension of SCDMs.
Novel features versus prior work (SCIM/SCG/MAID)
- Explicit decision constraints (Γ) and open root variables.
- Use of deterministic structural functions for endogenous non-decision nodes.
- Composability: SCDMs can be composed/decomposed; SCDPs inherit this and allow sequential decomposition in many dynamic problems.
Expressiveness and modeling benefits
- SCDPs strictly exceed the expressiveness of MDPs and POMDPs because they do not require rational (Bayesian) belief formation and can endogenize memory-formation and cognitive-resource tradeoffs—enabling modeling of resource-rational agents.
- Variable discounting: SCDPs can represent quasi-hyperbolic and state-dependent discount processes used in macro and behavioral economics.
Computational implications
- Composability can improve solution efficiency: sequential decomposition (e.g., two-period consumption problems, habit-formation variants) reduces joint optimization into simpler subproblems and yields computational gains; the paper includes complexity analysis illustrating these gains.
Applications & tooling
- Proposed applications: policy simulation for digital economy, mechanism design in information systems, digital twin modeling of cyberinfrastructure and sociotechnical systems.
- Implementation underway: scikit-agent project (documentation and code provided by authors).

Data & Methods

Methodological approach: formal, theoretical development using Pearl-style structural causal graphs and influence-diagram ideas; formal definitions and theorems (SCDM, SCDP), constructive examples, and computational complexity analysis.
Illustrative examples:
- Two-period Fisher consumption problem recast as an SCDM to illustrate separation of state, decision, utility, structural equations, constraints, and sequential decomposition via a bridge variable (assets).
- Multi-period/dynamic examples including consumption + portfolio allocation and habit formation to demonstrate decomposition and solution efficiency.
Expressiveness proofs: arguments and constructions showing SCDPs can represent problems that POMDPs/MDPs cannot (e.g., endogenous memory-formation, non-Bayesian belief processes, variable discount processes).
Computational analysis: complexity comparisons showing how compositional decomposition reduces solution cost in special cases (e.g., sequential decomposition).
Implementation: reference to scikit-agent (library and docs) as an initial implementation platform; code and documentation provided by authors.

Implications for AI Economics

Modeling realism and policy simulation
- Enables richer, composable models of digital economies and socio-technical systems where agents, algorithms, and subsystems interact via causal data flows.
- Facilitates policy experiments where agent cognition, memory, and patience (discounting) vary endogenously.
Mechanism design & system design
- The explicit causal and constraint structure supports mechanism-design analysis for information systems (e.g., incentives, budget/binding constraints, reward-manipulation scenarios).
- Composability allows modular design and testing of subsystem designs (digital twins) and the combination of strategic actors with physics-informed components.
Resource-rational agents & alignment
- By endogenizing cognitive costs and memory formation, SCDPs provide a principled framework to model boundedly rational agents and tradeoffs in computational resources—relevant to AI alignment and safety when agents operate under resource limits.
Econometric and empirical modeling
- The graph- and structural-equation framing facilitates causal identification strategies and counterfactual reasoning in economic settings that involve algorithmic actors.
Practical next steps & limitations
- Current paper focuses on single-agent formalisms and theoretical properties; multi-agent extensions, learning/fitting algorithms, equilibrium computation, and empirical validation remain future work.
- Implementational and scalability challenges will depend on algorithmic development to exploit composability in large, realistic systems.

Summary: SCDMs/SCDPs offer a causal, composable, and expressive modeling class that bridges structural causal modeling and dynamic decision theory, enabling richer, modular representations of boundedly rational and variable-patience agents in AI-augmented economic systems.

Assessment

Paper Typetheoretical Evidence Strengthn/a — The paper develops formal causal–decision frameworks and theoretical results without empirical estimation or causal inference from data, so empirical evidence strength is not applicable. Methods Rigorhigh — The contribution is a formal methodological/theoretical development that defines new model classes (SCDMs, SCDPs), proves composition and expressiveness properties (including a strict comparison to POMDPs), and explicitly characterizes modeling features like endogenous memory and variable discounting; however, empirical and computational evaluation is limited or absent. SampleNo empirical sample or observational/experimental data; the paper provides formal model definitions, theoretical propositions/arguments, and (presumably) illustrative examples rather than dataset-driven analysis. Themesorg_design productivity human_ai_collab GeneralizabilityNo empirical validation on real-world AI or economic systems presented, so practical applicability is untested, Requires specification of causal structure/structural equations which may be unavailable or misspecified in many real settings, Computational tractability and scalability to large, high-dimensional systems are not demonstrated, Estimation/inference procedures for fitting SCDMs/SCDPs to data are not provided, limiting transfer to empirical research, Assumes agent decision constraints and discounting can be represented within the formalism; misspecification of these primitives could limit external validity

Claims (11)

Claim	Direction	Confidence	Outcome	Details
We present two new classes of causal models of decision-making agents: Structural Causal Decision Models (SCDMs) and Structural Causal Decision Processes (SCDPs). Research Productivity	positive	high	introduction of new model classes (SCDMs and SCDPs)	0.2
Structural Causal Decision Models (SCDMs) expand on Structural Causal Influence Models by explicitly representing the causal relationships between model variables and the payoffs of agent decisions. Research Productivity	positive	high	explicit representation of causal relationships between variables and payoffs	0.2
In SCDMs, agent decisions can be constrained by their causal antecedents (i.e., decisions can be constrained by their causal parents). Research Productivity	positive	high	decision constraints by causal antecedents	0.2
SCDMs can have open root variables for which no probability distribution or structural equation is given. Research Productivity	positive	high	support for open root variables in model formalism	0.2
SCDMs have a well-defined and computationally useful property of composability. Research Productivity	positive	high	composability of causal decision models	0.2
A Structural Causal Decision Process (SCDP) is defined as a recurring SCDM with a discount variable. Research Productivity	positive	high	definition of SCDP as recurring SCDM with discounting	0.2
SCDPs inherit the composition properties of SCDMs (i.e., SCDPs benefit from SCDM composability). Research Productivity	positive	high	inheritance of composability by SCDPs	0.12
SCDPs are strictly more expressive than POMDPs because they do not assume rational belief formation. Research Productivity	positive	high	expressiveness relative to POMDPs (ability to represent non-rational belief formation)	0.2
An SCDP can endogenously model the memory-formation process and is thus useful for modeling resource‑rational agents in dynamic settings. Skill Acquisition	positive	high	ability to model endogenous memory formation / resource-rational agents	0.12
SCDPs are capable of modeling variable discounting, a tool used widely in social scientific modeling. Research Productivity	positive	high	modeling of variable discounting	0.2
SCDPs are a useful framework for policy simulation for the digital economy, mechanism design for information systems, and digital twin modeling of cyberinfrastructure. Governance And Regulation	positive	high	usefulness for policy simulation, mechanism design, and digital twin modeling	0.02