Information Theoretic Bayesian Optimization over the Probability Simplex

This paper introduces $\alpha$-GaBO, a novel family of Bayesian optimization algorithms that leverages information geometry to construct Matérn kernels and geometric optimizers tailored for the probability simplex, demonstrating superior performance over constrained Euclidean approaches in optimizing mixtures and robotic control tasks.

The Gist

Imagine you are a master chef trying to create the perfect soup. You have a list of ingredients (salt, pepper, carrots, onions), but there's a catch: you must use exactly 100% of the bowl's capacity. If you add more salt, you must remove something else. Your goal is to find the exact mix that tastes the best, but tasting the soup is expensive, time-consuming, and you can't taste it perfectly every time (it's "noisy").

This is the problem of optimizing on a probability simplex. It's a fancy way of saying: "How do we find the best mix of things that must add up to 100%?"

This paper introduces a new, smarter way to solve this problem called $\alpha$-GaBO. Here is how it works, explained through simple analogies.

1. The Problem: The "Flat Map" vs. The "Curved Reality"

Most computer algorithms that try to find the best mix treat the ingredients like they are on a flat, square grid (Euclidean space). They think, "If I add a little salt, I just move a tiny bit to the right."

But the reality of mixing ingredients is different. It's more like a curved surface.

• The Old Way (BORIS/Standard BO): Imagine trying to navigate a curved mountain using a flat paper map. You might take a shortcut that looks good on the paper but leads you off a cliff or into a swamp. In math terms, this ignores the "geometry" of the problem, leading to slow or suboptimal results.
• The New Way ($\alpha$-GaBO): This paper says, "Let's stop using the flat map. Let's use a globe." It realizes that the space of all possible mixes is actually curved, like the surface of a sphere.

2. The Magic Trick: The "Sphere Map"

The authors use a clever mathematical trick called the Sphere Map.

• The Analogy: Imagine the probability simplex (your soup ingredients) is a flat, triangular piece of paper. It's hard to do complex math on a crumpled piece of paper.
• The Solution: They invent a magical projector that takes that flat triangle and stretches it perfectly onto the surface of a ball (a sphere).
• Why do this? Math on a sphere is well-understood and very efficient. It's like having a GPS that works perfectly on a globe, whereas your old GPS only worked on a flat map. By moving the problem to the sphere, they can use powerful, pre-existing tools to find the best soup recipe much faster.

3. The "Smart Compass": The $\alpha$-Connection

Once the problem is on the sphere, the algorithm needs to know which direction to walk to find the best soup. In the old days, you just walked in a straight line. But on a sphere, "straight lines" are curves (like flight paths).

The paper introduces a family of "Smart Compasses" controlled by a dial called $\alpha$.

• $\alpha = 0$ (The Balanced Compass): This is the standard, most reliable way to walk on the sphere. It treats the geometry perfectly, ensuring you don't get lost. It's great for finding the best mix even if the answer is right on the edge of the bowl (e.g., a soup with only salt and no water).
• $\alpha = -1$ (The Exponential Compass): This is a specialized compass that is very good at exploring the middle of the bowl but struggles if the best answer is right on the very edge. It's like a compass that works great in the open ocean but gets confused near the shore.

By having this dial, the algorithm can choose the best "walking style" depending on the specific problem.

4. Real-World Results: From Soup to Robots

The authors tested this new method on three very different real-world problems:

1. Chemical Mixtures (The Soup): They tried to find the best mix of chemicals for solar cells. The new method found better mixes faster than the old methods.
2. Robot Classifiers (The Jury): Imagine a robot that needs to decide how to move by listening to a "jury" of 8 different simple decision-makers. The robot needs to figure out the perfect voting weight for each juror. The new method found the perfect jury mix faster.
3. Robot Control (The Dance): They taught a humanoid robot to walk around a pillar without falling or hitting it. The robot had to balance multiple tasks (move left hand, move right hand, stay upright, avoid pillar). The new method helped the robot learn the perfect "dance" of task priorities much faster and more reliably than before.

The Bottom Line

$\alpha$-GaBO is like upgrading from a flat, 2D map to a 3D globe for navigating complex mixtures.

• Old way: "Let's guess and check on a flat grid." (Slow, sometimes gets stuck).
• New way: "Let's project the problem onto a sphere, use a specialized compass, and glide to the solution." (Fast, efficient, and finds better answers).

This is a big deal for anyone trying to optimize mixtures, from designing new medicines and materials to teaching robots how to move gracefully.

Technical Summary

Here is a detailed technical summary of the paper "Information Theoretic Bayesian Optimization over the Probability Simplex" by Pavesi, Candelieri, and Jaquier.

1. Problem Statement

Bayesian Optimization (BO) is a powerful technique for optimizing expensive, black-box functions. However, many real-world applications involve optimizing probabilities and mixtures, where the search space is the probability simplex ($\Delta_d$). The simplex is a constrained, non-Euclidean domain defined by vectors with non-negative entries summing to one.

Existing approaches face significant limitations:

• Constrained Euclidean BO: Treating the simplex as a Euclidean space with constraints ignores the intrinsic geometry of the domain, often leading to suboptimal performance.
• Previous Geometry-Aware Attempts (e.g., BORIS): While BORIS attempted to use the Wasserstein distance, its practical implementation approximated this distance with the Euclidean norm, effectively reverting to a constrained Euclidean approach. Furthermore, Wasserstein geometry lacks closed-form Riemannian gradients, hindering efficient optimization of the acquisition function.

There is a lack of a rigorous, geometry-aware BO framework specifically tailored to the probability simplex that leverages its information-theoretic structure.

2. Methodology: $\alpha$-GaBO

The authors propose $\alpha$-GaBO, a family of geometry-aware Bayesian optimization algorithms grounded in Information Geometry. The methodology consists of two main pillars:

A. Geometric Foundation: Isometry to the Sphere

The authors leverage the Fisher-Rao metric to establish an isometry (a distance-preserving map) between the probability simplex $\Delta_d$ and the positive orthant of a unit hypersphere $S^d_{\geq 0}$.

• Sphere Map ($\phi$): Defined as $\phi(x) = 2\sqrt{x}$ (element-wise square root). This map transforms the simplex into a subset of the sphere.
• Benefit: This allows the authors to utilize well-established Matérn kernels defined on Riemannian manifolds (specifically the sphere) and pull them back to the simplex. This solves the challenge of defining valid positive-definite kernels on manifolds with boundaries.

B. The $\alpha$-Connection Family

To optimize the acquisition function on the manifold, the authors utilize the $\alpha$-connection from information geometry. This is a one-parameter family of connections ($\nabla^{(\alpha)}$) that interpolates between the mixture connection ($\alpha=1$) and the exponential connection ($\alpha=-1$). The Levi-Civita connection corresponds to $\alpha=0$.

The framework offers two specific algorithms based on parameter choices:

1. $\alpha_{-1}$-GaBO (Exponential Connection):
   • Uses the exponential connection ($\alpha = -1$).
   • The exponential map is unconstrained on the interior of the simplex, mapping the entire tangent space to the interior.
   • Limitation: It cannot reach the boundary of the simplex (vertices/faces) as it requires infinite tangent vectors to approach the border. It is unsuitable if the optimum lies on the boundary.
2. $\alpha_{0}$-GaBO (Levi-Civita Connection):
   • Uses the Levi-Civita connection ($\alpha = 0$).
   • The exponential map allows tangent vectors to reach the boundary of the simplex.
   • Implementation: It effectively performs optimization on the sphere (via the sphere map) and maps results back to the simplex. This provides a closed-form expression for Riemannian operations and handles boundary solutions naturally.

3. Key Contributions

• Novel Framework: Introduction of $\alpha$-GaBO, the first rigorous geometry-aware BO framework specifically designed for the probability simplex using information geometry.
• Kernel Construction: Derivation of valid Matérn kernels on the simplex by exploiting the isometry with the hypersphere, overcoming the lack of general kernel constructions for manifolds with boundaries.
• Flexible Optimizers: Development of a one-parameter family of acquisition function optimizers ($\alpha$-connections) that allow users to incorporate prior knowledge about the problem geometry (e.g., whether the optimum is likely to be on the boundary).
• Closed-Form Operations: Provision of closed-form expressions for Riemannian gradients, Hessians, and exponential maps for specific $\alpha$ values, enabling efficient implementation.

4. Experimental Results

The authors evaluated $\alpha$-GaBO on synthetic benchmarks and three real-world applications, comparing it against constrained Euclidean BO and BORIS.

• Benchmark Functions (Ackley, Rosenbrock, Griewank):
  • $\alpha$-GaBO models consistently matched or outperformed constrained Euclidean approaches, converging to lower function values with lower variance.
  • Performance gains were most significant in lower dimensions ($d=2, 5$).
• Optimal Mixtures of Components:
  • Concrete Strength: The optimum was found on the boundary. $\alpha_0$-GaBO and Euclidean BO on the sphere performed well, while $\alpha_{-1}$-GaBO struggled (as expected, due to its inability to reach the boundary).
  • Chemical Mixtures (Olympus): $\alpha$-GaBO models showed lower function values and significantly lower variance compared to BORIS and Euclidean BO, indicating more consistent performance.
• Mixture of Classifiers:
  • On a robot navigation task, all models performed similarly, though $\alpha_{-1}$-GaBO and Euclidean BO on the sphere converged slightly faster.
• Robotic Multi-Task Control:
  • In a complex task involving a humanoid robot avoiding obstacles while reaching targets, $\alpha_0$-GaBO performed best overall, converging faster to lower loss values with significantly lower variance than Euclidean counterparts. The robot successfully generated collision-free trajectories.

5. Significance and Conclusion

This paper bridges the gap between information geometry and Bayesian optimization. By rigorously applying the Fisher-Rao metric and $\alpha$-connections, the authors demonstrate that respecting the intrinsic geometry of the probability simplex leads to superior optimization performance compared to treating it as a constrained Euclidean space.

Key Implications:

• Robustness: Geometry-aware methods provide more consistent results (lower variance) across different seeds and problem types.
• Boundary Handling: The choice of $\alpha$ allows practitioners to tailor the algorithm to the specific nature of their problem (interior vs. boundary optima).
• Future Directions: The authors suggest this framework could extend to categorical data (via flow matching relaxations) and other information manifolds like symmetric positive-definite matrices.

In summary, $\alpha$-GaBO provides a mathematically sound and practically superior alternative for optimizing mixtures and probabilities, outperforming existing state-of-the-art methods in data efficiency and solution quality.