Mathematical Optimization

ORACLE: Near-Optimal Exploration as Certified Set Approximation

A note on ORACLE, which turns near-optimal energy-system exploration from point generation into an inner/outer approximation problem with a certified distance metric.

In fact, this paper is a very sophisticated method for finding the convex near-optimal solution space of a linear-programming energy system. Its emphasis is set approximation rather than stochastic or nonlinear optimization, but it fits naturally within the broader Mathematical Optimization category. I am writing this review because the paper contains ideas that are too original, precise, and frankly lovely not to record and share.

In energy-system planning, the single cost-optimal solution is often too thin a basis for decision-making. It tells us which design is cheapest under the model’s objective and constraints, but it does not tell us what else remains possible when the planner accepts a small cost increase. That surrounding near-optimal set is where many real decisions are made: one design may be slightly more expensive but easier to permit, less exposed to supply-chain risk, more politically acceptable, or better aligned with a technology preference that the model did not encode. The need is therefore not just to find alternative points, but to understand the shape of the acceptable design space. ORACLE’s contribution begins there. It does not propose a fundamentally new energy-system model; it changes near-optimal exploration from point generation into a geometry problem over the convex set of acceptable designs.

Classical Modelling to Generate Alternatives (MGA) usually operationalizes this need through repeated directional searches. Each run selects a weight vector or diversity criterion and returns one more feasible near-optimal design. This is useful, but the object produced by the algorithm is still a finite set of sampled points. ORACLE changes the object being computed. It maintains an inner approximation that is guaranteed feasible and an outer approximation that is guaranteed to contain the true region, then uses the largest distance between the two as both an error certificate and the next exploration target. The distinction is therefore not cosmetic: MGA is centered on solution generation, while ORACLE is centered on set approximation with a computable convergence metric.

The Region Being Approximated

Start from a cost-optimal LP,

v* = minx c⊤x subject to Ax≤b, Fx=d.

The full vector x may contain capacity, dispatch, storage operation, import/export, and many other decisions. But the analyst usually wants to understand only a smaller set of design variables,

z=Sx.

The near-optimal region in this projected design space is

Zε = { z: ∃x, Sx=z, Ax≤b, Fx=d, c⊤x ≤ v*(1+ε) } .

This set matters because a policy maker or system planner rarely cares only about the unique cheapest solution. A design that is 5% or 10% more expensive may be preferred because of political feasibility, social acceptance, supply-chain risk, regional fairness, or technology preference. These factors are often not explicit in the LP, but they matter after the optimization result leaves the model.

Why Ordinary MGA Is Not Enough

Standard MGA repeatedly solves problems of the form

minx,z wk⊤z subject to Sx=z, Ax≤b, Fx=d, c⊤x ≤ v*(1+ε).

Each iteration chooses a direction wk and finds an extreme near-optimal design in that direction. Random MGA, VMM, HSJ, SPORES, ERG, and Manhattan-style variants differ in how they choose the directions or encourage diversity. But the common weakness remains: they generate points, not a certified approximation of the set.

A point cloud can look diverse and still miss a large part of Zε. Worse, the user does not know how much was missed.

MAA moves closer to a region-based method by using a convex hull, but explicit convex-hull facet computation becomes expensive in moderate dimension. ORACLE avoids this by using the convex hull in convex-combination form rather than explicit halfspace form.

The Core ORACLE Invariant

ORACLE maintains a sandwich:

Ik ⊆ Zε ⊆ Ok .

The inner approximation Ik is the convex hull of known near-optimal points:

Ik = { Zkλ : 1⊤λ=1, λ≥0 } ,

where the columns of Zk are the near-optimal points already found. The outer approximation Ok is an intersection of valid halfspaces that must contain the true region.

The distance between the two approximations is

dk = maxzO∈Ok minzI∈Ik ‖zO-zI‖ ∞ .

This is the clever convergence metric. If the exploratory variables are capacity variables measured in GW, then dk=0.1 GW means the worst remaining coordinate-wise approximation error is at most 0.1 GW. That is much more interpretable than a volume gap.

Step 2: Find the Most Suspicious Outer Point

Step 2 asks, “Where is the current outer approximation farthest from what we have already certified?” Using the convex-combination form of the inner hull, the abstract problem is

dk = maxzO∈Ok minλ≥0,1⊤λ=1 ‖zO-Zkλ‖ ∞ .

The order of max and min is essential. For each candidate outer point zO, the lower problem computes its true shortest distance to the current inner hull. Then the upper problem chooses the outer point whose shortest distance is largest.

For a fixed zO, the lower-level distance can be written as an LP:

minλ,ρ ρ subject to -ρ1 ≤ zO-Zkλ ≤ ρ1, 1⊤λ=1, λ≥0.

Here ρ is the largest coordinate-wise difference between the trial point and the closest convex combination of known feasible points. It is the maximum of the PV gap, wind gap, gas gap, and so on.

One must not simply maximize ρ with these inequalities in a single flat problem, because then ρ could be inflated artificially. It has to be the optimal value of the lower-level distance problem. The paper handles this by reformulating the bilevel LP with LP optimality conditions; with the ℓ∞ norm, the final Step 2 problem becomes a single-level MILP.

The output of Step 2 is zO*, the most suspicious point in the current outer approximation. It may be truly near-optimal. It may also be an artifact of an outer approximation that is too loose. Step 2 does not know which case holds.

Step 3: Project the Trial Point Back to the True Model

Step 3 asks a different question: “Is there an actual near-optimal energy-system design close to this aggressive trial point?” It solves the projection problem

zf* ∈ argminzf∈Zε ‖zO*-zf‖ ∞ .

Because membership in Zε is defined through the original system model, the projection is written with the full variable x:

minx,ρ ρ subject to -ρ1 ≤ zO*-Sx ≤ ρ1, Ax≤b, Fx=d, c⊤x ≤ v*(1+ε).

This is much easier to interpret than Step 2. Step 2 says, “go to the place where the approximation is most empty.” Step 3 says, “now ask the real model where the closest feasible point actually is.”

If the optimal distance is zero, then zO* itself is a true near-optimal point and can be added to the inner approximation. If the distance is positive, the trial point is infeasible, but the projection zf* is still a valid near-optimal point and can be added to the inner approximation. In the infeasible case, the dual information from the projection can also generate a separating hyperplane that removes zO* from the outer approximation while preserving the true near-optimal region.

That is the loop:

  1. Step 2 chooses the worst unexplored outer point.
  2. Step 3 checks it against the original model.
  3. The inner approximation grows by adding a real near-optimal point.
  4. If the trial point was infeasible, the outer approximation shrinks by a valid cut.

The monotonic structure is the reason the method feels so clean:

Ik ⊆ Ik+1 ⊆ Zε ⊆ Ok+1 ⊆ Ok .

What Makes the Idea Original

The paper’s originality is in combining three choices.

First, it uses the convex hull of known near-optimal points in convex-combination form, not explicit facet form. This is the difference between asking for a tractable LP membership check and asking a high-dimensional convex-hull algorithm to expose every face of the polytope.

Second, it defines the gap between inner and outer approximations as an interpretable max-min distance. The metric has the same unit as the design variables, so convergence can be stated in the language of capacity, not only in abstract polytope volume.

Third, it uses the original system model as an oracle. A trial point is projected back to the true near-optimal region; the projected feasible point grows the inner hull, and an infeasible trial point produces information for cutting the outer approximation.

This is why ORACLE is more than a smarter MGA heuristic. It is a certified set-approximation scheme for convex near-optimal regions.

What Is Actually Guaranteed

The guarantees are strong, but only under the right assumptions.

If the near-optimal region is convex, then the convex hull of discovered near-optimal points is a valid inner approximation. If every outer cut is valid for the true region, then the outer approximation remains valid. If every point in the outer approximation is within tolerance of the inner approximation, then every point in the true region is also within that tolerance of the inner approximation.

This is the part that ordinary MGA does not provide. MGA can produce many interesting points, but it usually cannot say that the maximum unexplored error is below a specified threshold.

The limitation is just as important. The argument depends on convexity. Many realistic energy-system models include unit commitment, binary investment, modular technologies, minimum-load constraints, startup and shutdown costs, nonlinear efficiency, or degradation effects. In those settings, the near-optimal region may be nonconvex, and a convex combination of two feasible designs need not be feasible. ORACLE is therefore most naturally suited to continuous LP, convex, or quasiconvex planning models.

That is why I would not classify this paper as a core nonlinear-optimization paper. But as an algorithmic idea, it is unusually elegant: it turns “generate diverse alternatives” into “approximate the feasible alternative landscape with an error certificate.”

References

Turan, E. M.; Moret, S.; Bardow, A. “ORACLE: A rigorous metric and method to explore all near-optimal designs for energy systems.” arXiv preprint arXiv:2509.26452, 2025. https://arxiv.org/abs/2509.26452