AEC Model: Agent, Environment & Community

• AEC Model is a conceptual framework that applies descriptive, predictive, and prescriptive analytics to integrate agents, environments, and communities.
• The model employs mathematical formulations and machine learning tools, such as ARIMA and random forests, to optimize decision-making under uncertainty.
• Its applications in supply chains, digital twins, and healthcare planning demonstrate its effectiveness in reducing costs and enhancing system performance.

The Agent–Environment–Community (AEC) model is not a formalized framework by that name in the referenced arXiv literature, but the closely aligned Descriptive–Predictive–Prescriptive (DPP) paradigm serves as the rigorous foundation for analytics workflows that encompass agent-level decision-making (agent), environmental data and forecasting (environment), and system-wide or policy-based coordination (community). This article itemizes the theoretical constructs, methodologies, mathematical underpinnings, and cross-sector applications of the DPP approach, which effectively operationalizes the AEC concept for technical audiences.

1. Conceptual Decomposition: DPP as Agent–Environment–Community

The DPP pipeline explicitly distinguishes three analytic roles directly mapping to the AEC components:

• Descriptive (Environment): Aggregates and processes environmental data—observed features, outcomes, system states—synthesizing the empirical substrate upon which all subsequent modeling depends. For example, in a mental health workforce application, this involves collating NHS, population, and labor data series (Hewage et al., 2024).
• Predictive (Agent in Environment): Uses the environment's data to model conditional probability distributions, forecasts, or process dynamics influencing agent decisions. This may leverage machine learning algorithms (e.g., ARIMA, Gradient-Boosted Trees, LSTM) to project future demand, asset states, or risk (&&&1&&&, Hewage et al., 2024).
• Prescriptive (Community/Policy): Converts predictive outputs into actionable decisions, aligning agent or community actions with system-optimal outcomes through formal optimization, simulation, or policy analysis. Prescriptive analytics can yield local, agent-optimal actions or coordinated, system-level interventions (Bertsimas et al., 2014, Mortaz et al., 2021, Hewage et al., 2024).

This decomposition reflects the operational core of advanced analytics in domains spanning supply chain management, digital twins, and healthcare planning.

2. Mathematical Formalism Underpinning Agent–Environment–Community Workflows

The DPP/AEC workflow formalizes the decision-making process as a conditional stochastic optimization problem, typically using the following notation (Bertsimas et al., 2014):

• $x\in \mathcal X$: Decision vector (agent action).
• $Z\in\mathcal Z$: Environmental features.
• $Y\in\mathcal Y$: Uncertain outcome.
• $c(x;Y)$: Agent or system cost given action $x$ and outcome $Y$.

Oracle Program:

$$x^*(z) = \arg\min_{x\in\mathcal X} \mathbb{E}_{Y\sim F_{Y|Z=z}}[c(x; Y)]$$

Empirical Policy:

$$\hat x_n(z) = \arg\min_{x\in\mathcal X} \sum_{i=1}^n w_{n,i}(z) c(x; Y_i)$$

Weights $w_{n,i}(z)$, derived from ML models (kernel regression, CART, random forests), capture the impact of environmental context on optimal agent actions.

In cases of decision-dependent uncertainty—where actions influence environmental distributions—the formulation extends to optimizing $x$ given $F_{Y|Z, x}$ (Bertsimas et al., 2014):

$$\min_{x\in\mathcal X} \mathbb{E}_{Y\sim F_{Y|Z=z, x}}[c(x; Y)]$$

System-level (community) policies rely on stock-flow models, scenario trees, or multi-agent optimization to coordinate agents with respect to collective goals and constraints (e.g., healthcare workforce balance, fleet maintenance scheduling) (Hewage et al., 2024, Stadtmann et al., 2023).

3. Classification and Taxonomy of Methods

A broad suite of techniques populate the DPP/AEC pipeline, classified as:

DPP Role	Methods (examples)	Scope
Descriptive	Feature engineering, time series analysis, CCF, LASSO variable selection	Data curation, context
Predictive	ARIMA, ETS, LR, GBDT, DNN, LSTM, transfer learning, scenario modeling	Agent/environment
Prescriptive	Weighted empirical optimization, SAA, MINLP, MPC, stock-flow simulation	Policy/system action

Descriptive methods refine the representation of the environment. Predictive methods calibrate forecasts and quantifications conditional on environment state. Prescriptive methods solve for agent/system actions that minimize cost or maximize reward, often under uncertain, dynamic, or adversarial environmental response (Bertsimas et al., 2014, Hewage et al., 2024, Stadtmann et al., 2023).

4. Coupling Between AEC Levels: Train-Validate-Test and Prescription Loss

Traditional workflows decouple predictive and prescriptive modeling—training prediction models independently of their downstream usage in decision-making (agent actions) (Mortaz et al., 2021). The coupled validation approach replaces this paradigm by selecting model hyperparameters using prescriptive loss—the actual cost of the resulting system decision or agent action. The coupled workflow ensures that any bias introduced at the agent (prediction) stage is optimal for real-world outcomes, aligning agent and community interests under environmental uncertainty.

Generic coupled-validation step:

$$\hat\gamma^c = \arg\min_{\gamma\in\Gamma} \mathbb{E}_{(x, y)\in \text{Val}} [g(z^*(x; \hat\beta(\gamma), \gamma), y)]$$

where $z^*(x; \hat\beta(\gamma), \gamma)$ is the optimal decision given the predicted outcome (Mortaz et al., 2021).

Empirical results demonstrate that coupled-validation robustly reduces total system cost across synthetic and real datasets, especially under high uncertainty or misalignment between prediction and prescription objectives (Mortaz et al., 2021).

5. Community-Scale Optimization and Policy Analysis

Prescriptive/Community-level models extend agent–environment formalisms to system-wide coordination, using analytic constructs such as stock–flow models for healthcare workforce (Hewage et al., 2024) or mixed-integer programming for fleet maintenance (Stadtmann et al., 2023). Explicit scenario analysis, policy lever tuning, and regional customization allow for targeted interventions responding to local environmental and agent heterogeneity.

Example: The NHS nurse workforce planning framework decomposes joiner sources, applies multi-method forecasting, and supports policy scenario analysis with explicit equations for stock updates, joiner decomposition, and shortage quantification (Hewage et al., 2024):

• Stock update: $N_{r,t} = N_{r,t-1} + J_{r,t} - L_{r,t}$
• Shortage: $G_{r,t} = N_{r,t} - D_{r,t}$
• Joiner calculations for policy scenario mixing.

System-wide objectives typically target minimization of aggregate shortage or cost, with direct evaluation of policy efficacy via simulation or constrained optimization.

6. Applications Across Domains

The DPP/AEC architecture is realized in diverse contexts:

• Inventory and Supply Chain: Integration of auxiliary demand signals, random-forest-based prescription, evaluation by coefficient of prescriptiveness $P$ (analogue of $R^2$), with empirical $P \approx 0.88$ in real-world media inventory control (Bertsimas et al., 2014).
• Digital Twins: Evolution from descriptive twins (environmental/asset monitoring) to predictive and prescriptive twins (maintenance scheduling, operational point selection), employing NWP, DNN, LSTM, and mixed-integer optimization (Stadtmann et al., 2023).
• Healthcare Workforce Planning: Supervised and time series ML for multivariate demand/supply forecasting, regional and scenario-parameterized prescriptive simulation for staffing and recruitment strategy evaluation (Hewage et al., 2024).

A plausible implication is that the DPP/AEC structure is modular and extensible, enabling incorporation of advanced statistical/machine learning and optimization algorithms as improvements arise in each role.

7. Validation, Metrics, and Limitations

Validation of the DPP/AEC pipeline typically employs out-of-sample testing of prescriptive policies, using metrics such as:

• Coefficient of Prescriptiveness ($P$):

$$P = 1 - \frac{\text{Cost(data-driven policy)} - \text{Cost(oracle)}}{\text{Cost(clueless policy)} - \text{Cost(oracle)}}$$

A value $P=1$ indicates oracle-optimality, $P=0$ no better than baseline (Bertsimas et al., 2014).

• Empirical prescription cost compared between decoupled and coupled pipelines (Mortaz et al., 2021).
• System-level shortage/cost accumulated over regions/scenarios in policy models (Hewage et al., 2024).

Identified limitations include reliance on comprehensive, high-quality environmental data, risks of overfitting in forecast models (especially with limited mid/long-run data), and the necessity of domain expertise for feature engineering and policy extraction (Hewage et al., 2024, Stadtmann et al., 2023).

---

In summary, the Agent–Environment–Community architecture, operationalized as the Descriptive–Predictive–Prescriptive analytics paradigm, synthesizes environmental characterization, agent-level prediction, and community-scale prescription in a unified, mathematically rigorous, and application-agnostic workflow. This enables robust decision-making under uncertainty across technical domains, from supply chains to digital infrastructure and public health systems (Bertsimas et al., 2014, Mortaz et al., 2021, Stadtmann et al., 2023, Hewage et al., 2024).