gloss-opt 1.0.0

GLOSS: Global-Local-Unexplored Sampling Strategy for batch surrogate optimization in vast chemical search spaces

GLOSS

Global–Local–Unexplored Sampling Strategy — a multi-strategy batch recommender for surrogate-based optimization in vast chemical search spaces.

---

What it does

Standard batch Bayesian optimization (BO) picks all q candidates per round by greedily maximizing a single acquisition function. When the surrogate—fit on scarce data—has locked onto a secondary peak rather than the global optimum, the whole batch is wasted.

GLOSS decomposes each q-point batch across three complementary streams that share one surrogate:

Stream	Role	Selection
Global	Exploitation	UCB acquisition s·μ(x) + κ·σ(x)
Local	Refinement	BallTree neighborhood around current best x* (top-K truncation, K = 500 by default; O(K) per round)
Unexplored	Exploration	Maximizes geometric distance to observed points; uses no surrogate signal

The Unexplored stream is the operational answer to the overfitting trap: it forces every round to deposit data in regions the surrogate has not yet seen, so its blind spots get filled even when the μ/σ predictions are unreliable.

---

Install

pip install gloss-opt

The PyPI distribution is gloss-opt (the bare gloss name was already taken on PyPI), but the Python import name is still gloss:

from gloss import GLOSS

Optional extras:

pip install "gloss-opt[nn]"   # + torch (for NN surrogate)
pip install "gloss-opt[ml]"   # + xgboost, lightgbm
pip install "gloss-opt[all]"  # everything above

Install from source:

git clone https://github.com/zbc0315/gloss.git
cd gloss
pip install -e ".[all]"

Python 3.9+ required.

---

Quick start

import numpy as np
from gloss import GLOSS

# A candidate pool of 10,000 points in 5 dimensions
candidates = np.random.rand(10_000, 5)

g = GLOSS(
    space={"candidates": candidates},
    direction="maximize",
    ratio={"global_best": 4, "local_best": 2, "unexplored": 2},
    ucb_kappa=2.0,
    diversity_radius=0.02,
)

# Bootstrap with a few initial measurements
X_obs = candidates[np.random.choice(len(candidates), 8, replace=False)]
y_obs = my_oracle(X_obs)                       # your evaluation here

# Round-by-round recommendation
for _ in range(20):
    batch = g.recommend(X_obs, y_obs, n_points=8)
    y_new = my_oracle(batch)
    X_obs = np.vstack([X_obs, batch])
    y_obs = np.concatenate([y_obs, y_new])

See benchmarks/bench_main.py for end-to-end runnable examples on QM9, Buchwald–Hartwig and a virtual reaction surface.

---

Reproducing the benchmark

git clone https://github.com/zbc0315/gloss.git
cd gloss
pip install -e ".[all]"
python -m benchmarks.bench_main --study all

The benchmark compares GLOSS against UCB-BO, BO(EI), GA and Random on three chemistry datasets across 5 seeds × 20 rounds:

Dataset	n	Source
Buchwald–Hartwig	3,955	Experimental yields
QM9 HOMO–LUMO gap	100,000	DFT, 20 RDKit descriptors
Arrhenius-2D	10,000	Virtual reaction surface

Headline numbers on QM9-100k (5/5 seeds, mean t₉₅):

Algorithm	t₉₅ (rounds)	Reach 95%
GLOSS (4:2:2)	7.2	5/5
UCB-BO	16.6	3/5
BO(EI)	18.4	2/5

→ 2.31× / 2.56× speedup over the two BO variants.

---

Documentation

• Algorithm details, design decisions, and a per-stream walkthrough are in the paper (link to be added on submission).
• Per-class API: gloss/gloss.py (top-level GLOSS class), gloss/strategies/ (the three streams), gloss/surrogate/ (RF / GP / NN backends).
• Benchmark scripts: benchmarks/bench_*.py.

---

Citation

If you use GLOSS in your research, please cite:

@article{gloss2026,
  title  = {GLOSS: A Multi-Strategy Sampling Framework for Optimization in Vast Chemical Search Spaces},
  author = {Zhang, Baicheng and Zhang, Guoqing and Luo, Yi and Jiang, Jun and Zhu, Zhuoying},
  year   = {2026},
  note   = {Submitted}
}

---

License

MIT. See LICENSE.