https://ucl-lasp.github.io/tag/optimization/OptimizationHugo Blox Builder (https://hugoblox.com)en-usSun, 25 Jan 2026 00:00:00 +0000https://ucl-lasp.github.io/media/icon_hu488c70cfa50b07216f285734af4abcd1_22080_512x512_fill_lanczos_center_3.pnghttps://ucl-lasp.github.io/tag/optimization/https://ucl-lasp.github.io/project/optimization-agents/Sun, 25 Jan 2026 00:00:00 +0000https://ucl-lasp.github.io/project/optimization-agents/<h2 id="overview">Overview</h2> <p>Many industrial problems (routing, scheduling, circuit design) are NP-hard combinatorial optimization challenges. We investigate whether learning-based agents can &ldquo;outsmart&rdquo; or accelerate classical solvers.</p> <h2 id="active-projects">Active Projects</h2> <h3 id="1-neural-combinatorial-optimization">1. Neural Combinatorial Optimization</h3> <p><strong>Goal:</strong> Learning heuristics from data. <strong>Details:</strong> Instead of hand-crafting heuristics for every new problem, we train <strong>RL agents</strong> to learn construction and improvement heuristics automatically. We focus on graph-based problems where the agent learns to traverse the graph to build a valid solution.</p> <h3 id="2-generalizable-solvers">2. Generalizable Solvers</h3> <p><strong>Goal:</strong> Agents that generalize across problem sizes. <strong>Details:</strong> A major limitation of neural solvers is generalization. We are designing architectures (based on GNNs and attention) that allow an agent trained on small graphs (e.g., 20 nodes) to zero-shot generalize to large-scale instances (e.g., 1000 nodes) without retraining.</p> <h2 id="works-done">Works Done</h2>