Qualified Educational Capacity Planning under Heterogeneous Student Support Needs: A Synthetic Benchmark and Decision-Support Framework

Technical Summary: Qualified Educational Capacity Planning under Heterogeneous Student Support Needs

Problem Formulation
The paper addresses the planning of qualified educational staff capacity within a single institution. Unlike standard operations models that treat staff capacity as fixed headcount, this work models capacity as qualified staff time—time that is both available and possesses the specific preparation required for a support task (e.g., remediation, language support). The problem is characterized by three dynamic constraints:

1. Heterogeneity: Support demands are categorized, and staff must possess specific qualifications ($k$) for specific categories ($g$).
2. Dynamic Qualification: Preparation is a continuous state ($S_{w,k,t}$) subject to decay ($\delta$) if not practiced and requires training ($\alpha$) to reach a hard qualification threshold ($\theta$).
3. Capacity-Consuming Training: Training consumes the same scarce staff hours required for current student service. Furthermore, educational services are non-storable; unmet demand creates a backlog that cannot be "stored" for later fulfillment, and capacity qualified too late cannot retroactively serve past demand.

The decision problem involves allocating staff hours between direct service ($x_{service}$) and training ($x_{train}$) under uncertainty regarding demand surges, staff absences, and the arrival of new support categories. These disruptions can be announced (visible in the forecast window) or surprise (hidden until onset).

Methodology and Benchmark Design
The authors introduce a synthetic, seed-controlled benchmark and a decision-support framework rather than a deployed scheduler. The system is a stylized service model with:

• State: Backlog of unmet hours, staff availability, and continuous preparation levels.
• Actions: Allocation of service and training hours subject to hard eligibility constraints (only qualified staff can serve) and seat caps for training.
• Dynamics: Backlog evolves based on demand minus service; preparation levels decay and increase with training.
• Objective: Minimize the sum of unmet demand costs and training costs over a finite horizon.

The benchmark includes six scenario families: stationary control, new support categories (announced/surprise), staff absences (announced/surprise), and demand surges. It employs exact feasibility discipline, ensuring all policies generate valid actions, and uses replay checksums for bit-for-bit reproducibility.

Policy Classes
The study compares several policy classes, explicitly declaring their information sets:

1. ServiceOnly: Allocates service proportionally to demand; never trains.
2. ReactiveGap: Trains only after uncovered demand exceeds a threshold.
3. StaticInsurance: Executes a fixed, lean cross-qualification plan from $t=0$ based on initial configuration, spending a fixed budget fraction.
4. WaterFillingTraining: Spreads a fixed training budget across observed gaps.
5. ForecastAwareMPC (Primary): A rolling-horizon Mixed-Integer Linear Programming (MILP) controller that optimizes service and training over a finite horizon ($H=6$), re-planning at each step. It treats surprise shocks as masked until they occur.
6. OracleUpperReference: A perfect-foresight peer (same MPC structure but with surprise shocks unmasked) used to isolate the value of foresight, not as a global upper bound.

Key Results and Regime Map
The central finding is not a universal "best" policy, but a regime map governed by the relationship between the training lag (time to acquire a new qualification) and the controller's reaction reach (planning horizon + recoverable backlog window).

1. Reaction-Feasible Regime: When a new qualification can be acquired within the controller's reaction reach (e.g., standard training rates, warm starts, or short disruption windows), the closed-loop MPC controller dominates.

   • In a frozen adversarial suite of 2,620 episodes (covering short windows, slow training, low slack, and shock-focused static plans), the controller won all 158 paired cells against static baselines.
   • Attribution analysis shows the value comes from just-in-time qualification acquisition, not just better service allocation.
   • The controller performs nearly identically to the perfect-foresight Oracle on announced shocks, confirming its legitimacy.

2. Pre-Positioning Regime: When the training lag exceeds the controller's horizon (e.g., a "cold start" with deep retraining requiring ~60 staff-hours, exceeding the $H=6$ horizon), lean static insurance wins by up to 2× on long windows.

   • This win is structural: the perfect-foresight Oracle (also limited to $H=6$) ties the primary controller and loses identically to the static plan, proving the failure is due to the horizon constraint, not the controller's logic.
   • Reactive trainers that start after onset are the worst performers, wasting effort on training that cannot qualify in time.

3. Structural Insufficiency: When total demand structurally outruns qualified capacity ($\rho < 1$), all policies collapse, and policy choice becomes irrelevant.

4. Backlog Perishability: Introducing a deadline (perishable backlog) shifts the boundary between regimes but does not erase them. The controller's advantage in the reaction-feasible regime does not rely on forgiving late service.

Significance and Claims
The paper claims to provide the first reproducible synthetic benchmark for qualified educational capacity planning that integrates dynamic skill states, training opportunity costs, and heterogeneous demand.

• Contribution: It establishes a regime map distinguishing when dynamic, just-in-time control is superior versus when static, pre-positioned insurance is structurally necessary.
• Contrast with Manufacturing: Unlike companion manufacturing studies where static insurance often wins near capacity boundaries, this educational model shows that dynamic control dominates in most realistic scenarios unless the training lag is structurally longer than the planning horizon.
• Limitations: The authors explicitly state that all results are stylized computational evidence on synthetic data. The framework makes no claims about real student outcomes, legal compliance, or individual placements. It serves as a testbed for studying decision problems, not a deployment tool.
• Tooling: The paper introduces EduCapacity Studio, a transparent interface for scenario analysis and bit-for-bit replay, designed as a reproducibility instrument rather than a recommendation engine.

In summary, the paper argues that the "best" strategy depends entirely on the ratio of training lag to the controller's reaction reach. If the lag is short, dynamic control wins; if the lag is long (making reaction structurally too late), lean static pre-positioning is the optimal structural response.