🤖 machine learning

A Finite Expression Method for Solving High-Dimensional Committor Problems

This paper proposes using the Finite Expression Method (FEX), which leverages reinforcement learning to discover optimal algebraic expressions, as an effective and interpretable tool for solving high-dimensional committor problems in Transition Path Theory, offering superior accuracy and the ability to reduce problem dimensionality compared to neural network-based approaches.

Original authors: Zezheng Song, Maria K. Cameron, Haizhao Yang

Published 2026-01-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zezheng Song, Maria K. Cameron, Haizhao Yang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are standing in a vast, foggy mountain range with two distinct valleys: Valley A (where you started) and Valley B (your destination). The terrain is shaped by a "potential energy" map, meaning some paths are easy to walk, while others are steep and difficult. You want to know: If I start at this specific spot, what are the odds I will reach Valley B before I wander back to Valley A?

In the world of physics and chemistry, this question is called the Committor Problem. It helps scientists understand rare events, like how a protein folds into its correct shape or how a crystal breaks.

The paper you provided introduces a new, clever tool called the Finite Expression Method (FEX) to solve this problem, especially when the mountain range has so many dimensions (directions to move) that it becomes impossible to map using traditional tools.

Here is a breakdown of the paper's ideas using simple analogies:

1. The Problem: The "Curse of Dimensionality"

Think of a standard map as a 2D piece of paper. If you want to draw a map of a 10-dimensional mountain range, traditional methods (like Finite Element Methods) try to build a grid, like a giant 3D chessboard, but with 10 dimensions.

The Issue: As you add dimensions, the number of grid squares needed explodes. For a 10-dimensional problem, you would need more grid points than there are atoms in the universe. This is the "Curse of Dimensionality." It makes the calculation too expensive and slow.
The Old Solution: Scientists started using Neural Networks (AI). These are like black boxes that learn to guess the answer by looking at millions of data points. They are good, but they often just give you a "black box" formula that you can't read or understand.

2. The New Solution: FEX (The "Lego" Approach)

The authors propose a new method called FEX. Instead of using a black box AI, FEX tries to build the answer using a specific set of mathematical Lego bricks.

The Analogy: Imagine you are trying to describe a complex shape.
- Neural Networks are like a sculptor who chisels away at a block of stone until it looks right. You get the shape, but you don't know how they did it.
- FEX is like a child with a box of Lego bricks. The child is told, "Build a house using only these specific bricks (sine, square, plus, minus, etc.)." The child tries different combinations until the house looks perfect.
How it works: FEX builds a "tree" of math operations. It starts with variables (like $x_1, x_2$ ) and connects them with operators (like $+$ , $\sin$ , $\times$ ). It uses a smart "coach" (Reinforcement Learning) to tell the child which bricks to pick next.

3. The Superpower: Finding the "Secret Code"

The most exciting part of this paper isn't just that FEX solves the problem; it's how it solves it.

The Magic: When a Neural Network solves the problem, it gives you a messy, unreadable formula. But because FEX builds the solution from simple, readable math bricks, it often discovers the true structure of the answer.
The Result: In their tests, FEX realized, "Hey, even though this mountain has 10 dimensions, the answer only actually depends on one specific direction!"
- Once FEX identifies this, it can throw away the other 9 dimensions.
- This turns a super-hard 10D problem into a super-easy 1D problem that can be solved with perfect precision.
- Analogy: Imagine trying to find a needle in a 10-dimensional haystack. A Neural Network might just say, "It's somewhere in the middle." FEX looks at the haystack and says, "Actually, the needle is only moving up and down. Let's ignore the left-right and front-back directions entirely."

4. The Experiments: How Did It Do?

The authors tested FEX on several "benchmark" mountain ranges (mathematical problems) that are known to be difficult:

Double-Well Potential: A landscape with two valleys.
Concentric Spheres: Moving between two nested balls.
Rugged Landscapes: Very bumpy, complex terrain.

The Findings:

Accuracy: FEX performed just as well as, or better than, the best Neural Network solvers.
Understanding: Unlike Neural Networks, FEX produced formulas that humans could read and understand.
Dimension Reduction: In cases where the problem was actually simple (just depending on one or two variables), FEX found that simplicity automatically. It didn't just guess the answer; it figured out the rules of the game.

Summary

The paper claims that FEX is a powerful new tool for solving high-dimensional probability problems. It works by searching for a simple, readable mathematical formula (like a sentence made of math words) rather than a complex, unreadable AI model.

Its biggest advantage is that it can simplify the problem for you. If the answer only depends on a few variables, FEX finds that out, allowing scientists to solve the problem with extreme accuracy using traditional, reliable math methods, rather than relying on "black box" AI guesses.

Technical Summary: A Finite Expression Method for Solving High-Dimensional Committor Problems

Problem Statement
The paper addresses the computational challenge of solving the committor problem for high-dimensional stochastic systems, specifically the overdamped Langevin process. The committor function, $q(x)$ , represents the probability of a system hitting a metastable state $B$ before state $A$ starting from a point $x$ . Mathematically, $q(x)$ is the solution to a backward Kolmogorov equation subject to specific boundary conditions. While central to Transition Path Theory (TPT) for quantifying rare events, transition rates, and reaction channels, solving this equation numerically in high dimensions ( $d > 3$ ) is notoriously difficult due to the "curse of dimensionality." Traditional mesh-based methods (finite difference, finite element) become prohibitively expensive as the dimension increases. Existing machine learning approaches, primarily based on deep neural networks (DNNs), offer a mesh-free alternative but often function as "black boxes" that lack interpretability and may struggle to identify the underlying algebraic structure of the solution.

Methodology: The Finite Expression Method (FEX)
The authors propose adapting the Finite Expression Method (FEX), originally introduced by Liang and Yang (2022) for general PDEs, to the committor problem. Unlike DNNs which approximate solutions using continuous differentiable functions with fixed architectures, FEX seeks the solution as a symbolic mathematical expression composed of a fixed, finite number of nonlinear functions and binary arithmetic operations.

The core workflow involves:

Representation: The solution is represented as a binary tree of a predefined depth (typically 4–6). Each node in the tree corresponds to a unary (e.g., $\sin$ , $\exp$ , $\text{Id}$ ) or binary (e.g., $+$ , $\times$ , $\div$ ) operator. Leaf nodes contain input variables or constants.
Optimization Formulation: The committor problem is cast as a variational minimization of a loss functional (based on the Dirichlet energy weighted by the invariant density). The goal is to find the optimal sequence of operators and their associated numerical coefficients (scaling and bias parameters) that minimize this loss.
Mixed Combinatorial Optimization: Solving for the optimal tree structure is a mixed combinatorial and continuous optimization problem.
- Combinatorial Aspect: FEX employs a reinforcement learning (RL) controller (a neural network) to search for the optimal sequence of operators. The controller outputs probability distributions for selecting operators at each tree node.
- Continuous Aspect: Once an operator sequence is sampled, the numerical coefficients are optimized using gradient-based methods (Adam and BFGS).
- Training Loop: The controller is updated using policy gradient methods to maximize the expected "score" (inverse of the loss) of the generated expressions. A candidate pool of top-performing expressions is maintained and fine-tuned.
Handling Singularities: For problems involving singular behavior near boundaries (e.g., concentric spheres), the solution model is augmented with known fundamental solutions (e.g., $1/|x|^{d-2}$ ) multiplied by FEX trees to capture the correct asymptotic behavior.

Key Contributions and Results
The paper evaluates FEX on several high-dimensional benchmark problems, including double-well potentials with hyperplane and sublevel set boundaries, concentric spheres, rugged Mueller's potential, and butane. The results highlight two primary contributions:

Competitive Accuracy: On benchmark test problems, FEX achieves accuracy comparable to or better than state-of-the-art neural network-based solvers. For instance, in the double-well potential with sublevel set boundaries at $\beta=10.0$ , FEX achieved a relative $L_2$ error of $0.009$ compared to $0.021$ for the neural network method.
Structural Identification and Dimensional Reduction: The most significant finding is FEX's unique ability to correctly identify the algebraic structure of the solution. The method automatically determines which variables the committor depends on and which it does not.
- In the double-well potential with hyperplane boundaries, FEX correctly identified that the solution depends only on the first coordinate $x_1$ , despite the problem being defined in 10 dimensions.
- In the sublevel set boundary case, it identified the dependence on $x_1$ and the radial distance $r = \sqrt{\sum x_i^2}$ .
- This capability allows the high-dimensional problem to be reduced to a low-dimensional one (e.g., 1D or 2D). Once reduced, traditional high-accuracy methods (such as Chebyshev spectral methods or finite element methods) can be applied to achieve machine precision ( $10^{-16}$ ), a feat not possible with standard DNN approaches which do not explicitly reveal the solution's functional form.

Significance and Claims
The authors claim that the primary significance of this work lies in the interpretability and structural discovery capabilities of FEX. While other high-dimensional solvers (including DNNs and tensor train methods) may approximate the solution values, they generally fail to reveal the underlying algebraic dependencies or symmetries of the system. FEX bridges the gap between data-driven approximation and symbolic learning, enabling researchers to:

Reduce complex high-dimensional problems to tractable low-dimensional ones.
Obtain solutions with arbitrary accuracy by solving the reduced problem with classical numerical methods.
Gain physical insight into the transition mechanisms by analyzing the identified algebraic expressions.

The paper concludes that while FEX requires careful setup (choice of tree depth, operator sets, and training parameters), its ability to uncover the "true" algebraic structure of the committor function makes it a powerful tool for analyzing rare events in high-dimensional stochastic systems.