This research, presented at the Operational Research Society Simulation Workshop 2010 (SW10), investigates a critical question in simulation modelling: how do different simulation paradigms represent human behaviour, and do they yield meaningfully different results? The study specifically compares a traditional Discrete Event Simulation (DES) model with a hybrid model combining DES and Agent-Based Simulation (ABS) for modelling both reactive and proactive staff behaviour within a human-centric complex system—a women's wear fitting room in a UK department store.
The core aim was to evaluate the impact of modelling proactive behaviour (staff taking initiative) alongside reactive behaviour (staff responding to requests) on simulated system performance, and to determine if the more complex DES/ABS approach provided significantly different insights than a well-designed DES model.
The paper contextualizes its work within three major Operational Research (OR) simulation methods.
DES models a system as a sequence of events over time. The system state changes only at discrete points in time when an event occurs. It is process-centric, excellent for modelling queuing systems, resource allocation, and workflow. In human behaviour modelling, individuals are often represented as passive entities flowing through processes.
ABS models a system from the bottom-up, comprising autonomous, interacting agents. Each agent has its own rules, behaviours, and possibly goals. It is entity-centric, ideal for modelling heterogeneity, adaptation, learning, and complex interactions between individuals. It naturally captures proactive, goal-directed behaviour.
SDS focuses on aggregate-level feedback and stock-and-flow structures. It is suitable for strategic, high-level policy analysis but is noted as inappropriate for modelling individual-level heterogeneity and behaviour, which is the focus of this study.
The case study is the fitting room operation in a women's wear department of a top-ten UK retailer. The system involves customers arriving, queuing for a fitting room cubicle, trying on clothes, and staff assisting them. The research objective was to use simulation to determine the efficiency of new management policies by simulating staff behaviour.
The study builds on prior work (Majid et al., 2009) that modelled only reactive behaviour, extending it to a mixed reactive-proactive scenario.
The traditional DES model represented customers and staff as entities. Staff proactive behaviour was modelled using conditional logic and state variables within the process flow. For example, a "staff state" variable could trigger a "proactive queue management" sub-process if the queue length exceeded a threshold.
The hybrid model used a DES framework for the overall process flow (arrivals, queueing, resource usage) but implemented staff as autonomous agents. Each staff agent had a set of rules governing its behaviour, including decision-making logic for when to switch from a passive state to a proactive intervention state based on perceived environmental conditions (queue length, customer wait time).
Both models underwent standard verification (ensuring the model works as intended) and validation (ensuring it accurately represents the real system). A key validation technique employed was sensitivity analysis, testing how model outputs changed in response to variations in key parameters (e.g., rate of proactive intervention, staff numbers).
The study's most significant finding was that for the specific behaviours modelled, the traditional DES model and the DES/ABS hybrid model produced statistically similar output performance measures (e.g., average customer wait time, staff utilization, queue length).
Hypothesis: DES/ABS would show different performance due to richer agent interactions.
Finding: No statistically significant difference in key outputs between DES and DES/ABS for this case.
Implication: A well-structured DES model can effectively capture simple proactive rules.
The sensitivity analysis confirmed that both models responded similarly to changes in input parameters, strengthening the conclusion that their functional representation of the system's behaviour was equivalent for this scenario. The addition of proactive behaviour, in general, improved system performance metrics (reduced waits) in both models compared to the purely reactive baseline.
While the PDF abstract does not detail specific formulas, the modelling would involve standard queueing theory and probability distributions. A simplified representation of the proactive rule in both models could be:
Proactive Intervention Rule (Pseudo-Logic):
IF (Staff_State == "Idle" OR "Available") AND (Queue_Length > Threshold_L) AND (Random(0,1) < Probability_P) THEN
Initiate_Proactive_Action() // e.g., organize queue, assist waiting customers
Staff_State = "Proactive"
Duration = Sample_Distribution(Proactive_Time_Dist)
END IF
In DES, this is a conditional check within the staff process. In ABS, this rule is part of the staff agent's behavioural rule set, potentially evaluated continuously or at decision points. The core mathematical difference is not in the rule itself but in its enactment framework—centralized process flow vs. decentralized agent evaluation.
Performance metrics like average wait time ($W_q$) and system utilization ($\rho$) are calculated similarly in both models:
$W_q = \frac{1}{N} \sum_{i=1}^{N} (T_{i,service\,start} - T_{i,arrival})$
$\rho = \frac{\text{Total Busy Time of Staff}}{\text{Total Simulation Time}}$
Scenario: Modelling a hospital ward nurse's behaviour.
The 2010 study paved the way for more nuanced investigations. Future directions include:
Analyst Commentary: A Pragmatic Reality Check
Core Insight: This paper delivers a crucial, often overlooked truth in simulation: model complexity is not inherently virtuous. The DES/ABS hybrid, while academically fashionable for modelling human behaviour, failed to produce meaningfully different operational insights than a competently designed traditional DES model for this specific problem scope. The real value wasn't in the agent-based architecture, but in the explicit codification of proactive behavioural logic.
Logical Flow: The research follows a robust, classical OR methodology: define behaviour (reactive/proactive), select a relevant case (retail fitting room), build comparable models (DES vs. DES/ABS), run controlled experiments, and use statistical tests (likely t-tests or ANOVA) to compare outputs. Its strength is in this disciplined comparability, a step often missing in papers that champion one methodology over another.
Strengths & Flaws: The study's strength is its practical, evidence-based approach. It challenges the assumption that "more detailed" (ABS) is always "better." However, its flaw lies in the simplicity of the proactive behaviour modelled—simple threshold-based rules. As noted in later ABS literature, such as the work on cognitive architectures (e.g., ACT-R, SOAR) integrated with agents, the true power of ABS emerges with learning, adaptation, and complex social interactions, which were not tested here. The study compares a "smart DES" to a "simple ABS," potentially underestimating the latter's potential.
Actionable Insights: For practitioners: Start with DES. Before investing in the development and computational overhead of an ABS model, rigorously test if a well-thought-out DES model can capture the essential decision logic. Use sensitivity analysis to explore behavioural rules. Reserve ABS for problems where heterogeneity, adaptation, or emergent network effects are the core research questions, not just individual initiative. This aligns with the principle of parsimony—the simplest adequate model is often the best.