DUGE: The Optimization Algorithm That Doesn't Exist in Prior Art

Directional Uncertainty Gradient Exploration (DUGE) is 72Energy's novel optimization algorithm — and after extensive prior art searches, nothing like it exists in published literature or patent databases.

What Makes DUGE Different

Traditional building energy optimization faces a fundamental challenge: the solution landscape is riddled with local minima. Bayesian Optimization gets trapped. Reinforcement Learning requires millions of episodes. Simulated Annealing converges slowly. DUGE takes a fundamentally different approach by computing directional uncertainty gradients — navigating the parameter space by following where the model is least certain, not where the gradient points.

The Technical Innovation

DUGE uses a Bayesian ensemble of 10 neural networks to quantify prediction uncertainty at every point in the optimization space. The key insight is the adaptive learning rate mechanism:

• High uncertainty regions: Increase exploration (learning rate scales by β=2.0x)

• Low uncertainty regions: Increase exploitation (learning rate scales by γ=0.5x)

• Physics-guided tunneling: When the optimizer encounters infeasible regions (violating thermodynamic constraints), it tunnels through using physics-informed shortcuts rather than retreating

Validated Performance

Our automated test suite validates DUGE against three baseline optimizers on identical building optimization problems. The results, verified daily via GitHub Actions CI:

• Full optimization in under 3 seconds (vs. minutes for Bayesian Optimization)

• Escapes 90%+ of local minima that trap conventional optimizers

• Maintains physics constraint satisfaction throughout the optimization trajectory

• Converges to superior solutions compared to RL and Simulated Annealing baselines

DUGE represents one of 72Energy's five patent families — and the one with the highest assessed licensing value. For utilities and ESCOs operating at portfolio scale, it means real-time building optimization that actually finds the global optimum, not a convenient local one.