Introduction
A fundamental challenge in automated scientific discovery is balancing model accuracy with physical simplicity. While the Weierstrass approximation theorem guarantees that any continuous function can be modeled by a high-degree polynomial, such models often lack physical meaning. They may fit data points perfectly while violating dimensional consistency—for instance, by attempting to add incompatible units like meters and seconds. Early discovery systems like BACON and ABACUS struggled with this trade-off, often requiring complex heuristics or producing dimensionally inconsistent results. This paper presents COPER ("COmputer PERception"), a system that leverages dimensional analysis to discover simple, exact functional laws by searching for an optimal representation base.
Dimensional Analysis and the Pi-Theorem
COPER's methodology is rooted in the Pi-theorem of dimensional analysis. This theorem states that any physically meaningful formula can be expressed as a function of dimensionless monomials—groups of variables where units cancel out, such as the Reynolds number in fluid mechanics. By restricting its search to functions invariant under unit transformations, COPER ensures that every candidate formula is dimensionally consistent. This constraint significantly prunes the search space, as the system only considers "meaningful" functions that remain valid regardless of the specific units (e.g., meters vs. feet) used for measurement.
The core innovation of COPER is the "change of representation base." A dimensional base is a subset of variables that are dimensionally independent. The difficulty in finding a simple law often stems from an inappropriate choice of base. Just as a numerical pattern might be hidden in base 10 but obvious in base 9, a physical law may appear as a complex, high-degree polynomial in one dimensional base but as a simple linear equation in another.
The COPER Algorithm
The COPER algorithm follows a structured search process designed to prioritize simplicity:
- Initialization: Start with the simplest possible functional form, typically a first-degree (linear) polynomial.
- Base Iteration: Systematically iterate through every possible valid dimensional base available from the input variables.
- Monomial Construction: For each base, calculate the corresponding dimensionless monomials.
- Function Fitting: Perform a regression fit using the chosen functional form on these monomials.
- Evaluation: If the error is below a predefined threshold, the law is considered discovered.
- Complexity Escalation: If no suitable fit is found across all bases, increase the complexity of the functional form (e.g., moving from linear to quadratic) and repeat the process.
Results: Uniformly Accelerated Motion
The system was evaluated using synthetic data for uniformly accelerated motion, defined by the variables distance ($S$), initial velocity ($v_0$), acceleration ($a$), and time ($t$). The goal was to rediscover the classic kinematic law $S = v_0t + \frac{1}{2}at^2$.
As shown in the results, the choice of base is critical to the success of the discovery process. When using $(v_0, t)$ as the base, the system attempted a linear fit that resulted in a massive error of $4.78 \times 10^6$. However, when the system switched the base to $(v_0, a)$ or $(a, t)$, the error plummeted to near zero ($1.40 \times 10^{-2}$ and $1.84 \times 10^{-4}$, respectively).
By identifying the correct base, COPER transformed a seemingly complex quadratic problem into a simple linear fit that exactly matched the underlying physical law. This demonstrates that searching the "base space" first can be far more efficient than searching for high-degree curves in a fixed coordinate system.
Discussion and Limitations
While COPER is highly effective for discovering polynomial laws, it faces challenges with non-polynomial relationships. Laws involving logarithms or transcendental functions currently require high-degree polynomial approximations per the Weierstrass theorem, which results in formulas that lose their human-readable elegance. Additionally, while the system's use of the entire dataset for fitting provides some protection against noise, formal investigation into noise robustness is a priority for future research. Finally, as the number of variables increases, the exhaustive search of all possible bases becomes computationally expensive, necessitating the development of more advanced heuristics to navigate the base space.
Summary updated