Learning soft interventions in complex equilibrium systems

Michel Besserve, Bernhard Schölkopf

TL;DR

This paper introduces a differentiable framework for designing optimal "soft" interventions in complex systems with feedback loops, such as global economies. By combining Lie groups with deep equilibrium models, the authors provide a way to optimize interventions that can mitigate counterintuitive outcomes like the "rebound effect" (where efficiency gains lead to higher total consumption).

What Problem Does This Paper Address?

Most classical causal inference frameworks assume that systems are acyclic (Directed Acyclic Graphs, or DAGs). However, real-world systems—especially socio-economic and biological ones—are full of feedback loops. In these cyclic systems, intervening on one variable can have paradoxical effects that ripple through the network and return to the source.

A classic example is the rebound effect (Jevons Paradox): improving the energy efficiency of a device lowers the cost of using it, which can actually increase total energy demand. Designing policies to navigate these complexities is difficult because the "equilibrium" state of such a system is often defined by self-consistent equations that are analytically intractable. The authors aim to bridge the gap between causal theory and the practical optimization of these complex, cyclic equilibria.

Methods

The authors propose a framework based on three main pillars:

  1. Smooth Structural Causal Models (SSCM): They define a version of causal models where variables are related through smooth maps on manifolds, allowing for cycles and deterministic parameters.
  2. Lie Interventions: To make interventions differentiable and "soft" (meaning they modify the functional relationships between variables without destroying the underlying causal structure), the authors use Lie groups. For example, a multiplicative Lie intervention might rescale the demand for a specific economic sector.
  3. Deep Equilibrium Models (DEQ): They leverage the DEQ approach to optimize these interventions. Instead of unrolling a fixed number of iterations, they treat the equilibrium as an implicit function. This allows them to use automatic differentiation and backpropagation through the equilibrium point itself, using techniques like Anderson acceleration to find solutions efficiently.

A key innovation is the concept of Invariant Soft Interventions. This involves using an "auxiliary node" (like a tax policy) to keep a specific variable (like price) invariant during an intervention, effectively "compartmentalizing" the effects of a policy to prevent unwanted systemic rebounds.

Datasets

The paper evaluates the framework using both toy models and semi-synthetic experiments based on real-world economic data:

  • Exiobase 3: A massive multi-regional input-output (MRIO) dataset. The authors specifically extracted data for 200 economic sectors across 49 world regions for the years 2012–2018.
  • Pymrio library: Used to process the Exiobase data to estimate technical coefficient matrices ($A$) and footprint intensity matrices ($R$) for countries like France and Germany.

Key Insights and Ideas

  • Interventions as Group Actions: By framing interventions as Lie group actions, the authors turn a discrete "what-if" problem into a continuous optimization problem. This allows them to use standard gradient-based tools from the deep learning toolbox.
  • Solving the Rebound Effect: The paper mathematically demonstrates that you can "cancel out" undesired feedback (like the backfire scenario in energy) by learning a secondary, adaptive intervention on an auxiliary variable.
  • Compartmentalization: They prove that if you can enforce invariance on the "boundary" nodes of a subsystem, you can design interventions for that subsystem independently of the rest of the complex network. This is a powerful idea for modular policy design.

Key Results

  • Equilibrium Estimation: The authors showed that Anderson acceleration significantly outperforms standard forward iteration for solving high-dimensional economic equilibrium equations (up to 200 dimensions), which is crucial for making the optimization loop feasible.
  • Sustainability Trade-offs: In experiments on French and German economic models, the framework successfully mapped the "Pareto frontier" between reducing Greenhouse Gas (GHG) emissions and preserving employment. They found that Germany has less "room" than France to reduce emissions before seeing significant employment impacts, highlighting how the underlying structure of an economy dictates policy efficacy.
  • Preventing Backfire: In a 3-sector toy model, a standard intervention on energy efficiency led to a "backfire" where total energy demand increased. The invariant intervention (using an adaptive tax) successfully maintained price levels and achieved the intended reduction in energy demand.

Limitations and Future Work

  • Linearity Assumptions: While the intervention framework is general, the base economic models used (Leontief models) are largely linear. The authors note that moving toward non-linear Integrated Assessment Models (IAMs) is a necessary next step.
  • Known Parameters: The current framework assumes the system's parameters (the "graph" and the equations) are known. In many real-world scenarios, these must be estimated from noisy data.
  • Identifiability: A major open question is how to identify these cyclic models and their Lie interventions purely from observational or experimental data, rather than relying on pre-existing economic tables.

For more details, the code is available at github.com/mbesserve/lie-inter.

Summary updated

Abstract

Complex systems often contain feedback loops that can be described as cyclic causal models. Intervening in such systems may lead to counterintuitive effects, which cannot be inferred directly from the graph structure. After establishing a framework for differentiable soft interventions based on Lie groups, we take advantage of modern automatic differentiation techniques and their application to implicit functions in order to optimize interventions in cyclic causal models. We illustrate the use of this framework by investigating scenarios of transition to sustainable economies.
BibTeX
@inproceedings{besserve2022learning,
  author = {Michel Besserve and Bernhard Schölkopf},
  title = {{Learning soft interventions in complex equilibrium systems}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2022},
  url = {https://openreview.net/forum?id=Cjxe5kCtAe},
}
Figure