The Design and Composition of Structural Causal Decision Processes

Sebastian Benthall and Alan Lujan have released a new paper on arXiv introducing two classes of causal models for decision-making agents: Structural Causal Decision Models (SCDMs) and Structural Causal Decision Processes (SCDPs). The work is motivated by the economics of computing systems, which often feature subsystems with endogenous limits on cognitive resources and value discounting. SCDMs extend Structural Causal Influence Models (SCIMs) by explicitly representing causal relationships between variables and agent payoffs, while also allowing agent decisions to be constrained by their causal antecedents. A key innovation is the inclusion of open root variables—variables without a predefined probability distribution or structural equation—which makes the models more flexible. The authors prove that SCDMs are composable, meaning larger systems can be built from smaller models in a computationally useful way.

Building on SCDMs, the paper defines Structural Causal Decision Processes (SCDPs) as recurring SCDMs with a discount variable. SCDPs inherit composability and are strictly more expressive than Partially Observable Markov Decision Processes (POMDPs). Unlike POMDPs, SCDPs do not assume rational belief formation; they can endogenously model the memory-formation process, making them ideal for modeling resource-rational agents in dynamic environments. They also support variable discounting, a tool widely used in social sciences. The authors suggest applications in policy simulation for the digital economy, mechanism design for information systems, and digital twin modeling of cyberinfrastructure.