💻 computer science

Modeling Bounded Rationality in Drug Shortage Pharmacists Using Attention-Guided Dynamic Decomposition

This paper proposes and validates an attention-guided dynamic decomposition framework that models bounded rationality in hospital pharmacists by prioritizing cognitive effort on critical drug shortages, demonstrating that such satisficing strategies enable stable decision-making under uncertainty without requiring complete state reasoning.

Original authors: Yaniv Eliyahu Amiri, Noah Chicoine, Jacqueline Griffin, Stacy Marsella

Published 2026-05-15

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yaniv Eliyahu Amiri, Noah Chicoine, Jacqueline Griffin, Stacy Marsella

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the head chef in a massive, chaotic kitchen. You have hundreds of ingredients (drugs) to manage. Some are running low, some suppliers are unreliable, and you have to feed hundreds of people every day. If you run out of a key ingredient, patients (or diners) go hungry or get sick.

The problem is that you can't possibly check every single jar, can, and box on every shelf every single minute. There isn't enough time, and your brain can't hold that much information at once.

This paper is about how expert pharmacists (the head chefs) actually solve this problem, and how the authors built a computer program to mimic that smart, "bounded" way of thinking.

The Problem: Trying to Do Everything at Once

In the old way of thinking, computers tried to be "perfectly rational." They tried to look at every drug, calculate the best move for every one, and plan weeks ahead.

The Analogy: Imagine trying to read every single book in a library to decide which one to recommend to a customer. It would take forever, and you'd never get anything done. In the real world, this "perfect planning" is so slow and complex that the computer crashes or takes too long to help.

The Human Solution: The "Urgency Filter"

The researchers interviewed real pharmacists and found something interesting. Experts don't try to solve the whole puzzle at once. Instead, they use a mental filter.

The Analogy: Think of a lighthouse in a stormy sea. The lighthouse doesn't shine a beam on the entire ocean; it focuses its bright light on the specific rocks that are dangerous right now. The rest of the ocean is just "dark water" that gets a quick, lazy glance.
How it works: Pharmacists look at a few key signals:
- Runway: How many weeks of supply do we have left?
- Uncertainty: Is the supplier lying or being vague?
- History: Has this drug caused trouble before?
- Impact: Is this a life-saving drug?

If a drug hits a certain "urgency" level, the pharmacist focuses their brain power on it. If it's low urgency, they just keep an eye on it without deep thinking.

The Computer Models: Two New Agents

The authors built two computer "agents" (programs) to test if this "focus filter" works better than trying to do everything.

The Expert Agent: This is the "copycat." It was programmed with the exact rules the human pharmacists described in the interviews. It knows exactly which drugs to focus on based on the "urgency" signals.
The Learner Agent: This is the "student." It starts with no idea which drugs are important. It tries different things, makes mistakes, and learns over time which signals (like "supplier uncertainty" or "usage rates") actually help it avoid running out of stock. Eventually, it learns to focus on the right drugs, just like the human expert.

The Results: Fast and Stable

The researchers tested these agents in a simulated hospital environment with three different timeframes: short (3 weeks), medium (10 weeks), and long (52 weeks).

Speed: The "Focus" agents were much faster. In the long-term tests, the "perfect planner" (which tries to look at everything) became so slow it was practically useless. The "Focus" agents made decisions in seconds.
Success: Even though they ignored most of the drugs most of the time, they were just as good at preventing stockouts (running out of medicine) as the perfect planner.
The Lesson: The paper argues that the most important decision isn't "What action should I take?" but rather "Where should I spend my brain power?"

The Takeaway

The paper suggests that in high-stress, complex situations (like a hospital pharmacy), you don't need to be a super-computer that calculates everything. You just need to be smart about what you pay attention to.

By focusing only on the "hot" problems and letting the "cool" problems sit on the back burner, you can make stable, good decisions quickly without burning out your brain (or your computer). The "Learner" agent proved that a system can even teach itself to do this focus trick over time, getting better at spotting the real dangers without needing a human to program every single rule.

Technical Summary: Modeling Bounded Rationality in Drug Shortage Pharmacists Using Attention-Guided Dynamic Decomposition

Problem Definition
Hospital pharmacists managing drug shortages operate in a high-stakes, partially observable environment characterized by uncertain supply dynamics, noisy demand signals, and time pressure. The decision problem involves managing hundreds of drugs with uncertain inventory levels, fluctuating utilization rates, and unreliable supplier delivery estimates. Traditional approaches to this problem, such as Partially Observable Markov Decision Processes (POMDPs), attempt to optimize actions over the full state space. However, these methods are computationally intractable for realistic drug shortage settings due to the exponential growth of the state space. Furthermore, interviews with practicing pharmacists reveal that human experts do not perform global optimization; instead, they dynamically allocate cognitive effort to a small subset of urgent drugs while passively monitoring the remainder. Existing decision-support models fail to capture this selective attention mechanism or provide the explainability required by experts to understand priority formation.

Methodology
The authors propose a computational framework that reframes the decision problem from optimizing actions over a fixed state representation to controlling the allocation of computational resources across state variables. This approach is grounded in bounded rationality and Recognition-Primed Decision (RPD) theories, which suggest experts use pattern recognition to decompose problems into high-risk and low-risk subsets.

The framework introduces an Attention-Guided Planning mechanism that dynamically decomposes the set of drugs into two subsets:

High-Urgency Subset: A small set of drugs selected for deep, costly reasoning using an online planner.
Low-Urgency Subset: The complementary set of drugs subject to low-cost monitoring.

Selection is driven by an Urgency Score ( $U_i$ ) for each drug, calculated as a weighted sum of normalized components:
$U_i = U_{runway}(s_i) + \sum_{f \in F \setminus \{runway\}} \beta_f \cdot U_f(s_i, f)$
Where $U_{runway}$ represents the weeks of supply remaining, and other terms ( $U_f$ ) capture uncertainty in inventory/utilization, atypical utilization levels, clinical impact, and prior shortage history. The weights ( $\beta_f$ ) determine the contribution of each feature.

Two agents are developed to test this framework:

Expert Agent: Encodes attention weights derived directly from semi-structured interviews with pharmacists. These weights serve as a cognitive baseline, prioritizing drugs based on expert-identified signals like runway risk and supply uncertainty.
Learner Agent: Uses a REINFORCE-style update to adapt the attention weights ( $\beta$ ) over time. Instead of learning an action-selection policy, the learner adjusts the attention allocation weights based on observed rewards relative to a moving baseline. This allows the agent to learn which prioritization factors best support stable decision-making under uncertainty while maintaining an interpretable attention mechanism.

Experimental Setup
The models were evaluated in simulated weekly drug shortage scenarios with partial observability and stochastic dynamics. The simulations involved $n=19$ drugs and included three scenario sets with varying horizons:

Set 1: 3-week horizon (controlled disturbances).
Set 2: 10-week horizon (compounding effects and delayed consequences).
Set 3: 52-week horizon (long-term stability and stress testing).

Agents were compared against four baselines: a Random agent, a Greedy agent (prioritizing low runway), a Heuristic agent (using fixed pharmacist-inspired rules), and a Full POMDP agent (complete state online planning).

Key Results

Computational Efficiency: Attention-guided agents significantly reduced planning time compared to the Full POMDP baseline. In Set 1, planning time dropped from 0.59s to ~0.3s per week; in Set 2, it dropped from 2.11s to ~0.7s. This efficiency gain stems from restricting deep planning to a focused subset of high-urgency drugs.
Performance Stability: The Expert Agent outperformed static baselines (Greedy and Heuristic) across all scenario sets, particularly in long-horizon scenarios (Set 3). The Heuristic agent suffered 4 stockouts and a massive negative reward (-4.9M), whereas the Expert Agent maintained zero stockouts and achieved a total reward of 36,236.
Learning Capability: The Learner Agent achieved performance comparable to the Expert Agent across all sets (e.g., Set 3 reward: 36,448 vs. 36,236) while successfully adapting its attention weights. Analysis of the Learner's weight evolution showed a shift toward prioritizing usage and usage uncertainty over time, indicating that demand-side signals become critical for long-term stability.
Bounded Focus: Both agents maintained a bounded focus set (maximum of eight drugs) throughout the 52-week scenarios, demonstrating that stable decision-making can be achieved without exhaustive state reasoning.

Significance and Claims
The paper claims that for high-dimensional, partially observable decision problems, the primary decision is not what action to take, but where to allocate cognitive reasoning effort. The authors argue that global optimization strategies are not strictly necessary for strong performance in these domains. Instead, attention-guided, satisficing strategies can reduce problem complexity by dynamically decomposing the state space, thereby achieving tractability without sacrificing decision quality or stability.

The framework offers a computationally feasible alternative to full-state POMDPs that aligns with observed human expert behavior. It provides a mechanism for explainability by explicitly modeling how priorities are formed and updated. The authors note that while the model is simulation-based and uses simplified supply chain dynamics, it establishes a general cognitive principle: when a problem is too large to plan exhaustively, effective agents must explicitly manage the scope of their reasoning. Future work is identified as needing prospective validation with real operational data and the incorporation of richer social dynamics and inter-hospital coordination.