Technical Summary: Multi-Objective Optimization of Resource-Specific Handover Policies

Problem Definition

Efficient resource allocation is a critical challenge in business process management (BPM), directly impacting cost, throughput time, and utilization. While recent Reinforcement Learning (RL) approaches have demonstrated promise in deriving adaptive allocation policies, they typically rely on global policies based solely on task availability and resource capacity. These methods treat resources as passive recipients of assignments, neglecting the active decision-making role of resources (particularly humans) and the inter-resource collaboration patterns that shape real-world task handovers.

Existing optimization approaches often focus on assigning tasks to resources or determining staffing levels but fail to optimize resource-specific handover policies. A handover policy defines, for a specific resource completing a specific activity, which resource should perform the subsequent activity. The authors argue that ignoring these local, resource-specific dependencies overlooks a key leverage point for process improvement. The core problem addressed is the multi-objective optimization of these handover policies to find optimal trade-offs between conflicting objectives (e.g., cost vs. waiting time) while accounting for the collaborative structure of the process.

Methodology

The proposed approach combines a Multi-Agent System (MAS) based process simulator with a multi-objective evolutionary algorithm (NSGA-II) to discover Pareto-optimal, resource-specific handover policies.

1. Simulation Model

The foundation of the approach is a simulation model $M = (A, S, \pi_0, \beta)$, where:

• $A$: A set of agents representing resources, characterized by availability calendars and capabilities (activities they can perform and execution time distributions).
• $S$: Global simulation parameters (e.g., case interarrival distributions).
• $\pi_0$: The "as-is" handover policy.
• $\beta$: The control-flow policy determining the next activity.

This model can be manually specified or automatically discovered from event logs using tools like AgentSimulator.

2. Optimization Framework

The system employs a genetic algorithm to search for a set of handover policies $\pi$ that minimize a cost vector $c(L) = [f_1(L), \dots, f_k(L)]$, where $L$ is the event log generated by the simulation. The objectives are minimized simultaneously to identify the Pareto front.

The optimization workflow consists of three iterative steps:

1. Initial Population Creation: An initial population of policies is generated based on the "as-is" policy $\pi_0$. The search space is pruned by classifying decision points as either Fixed (non-negotiable business rules, excluded from optimization) or Variable (flexible parts). The population includes random policies for exploration, mutated versions of $\pi_0$ for exploitation, and a copy of $\pi_0$ as a baseline.
2. Simulation and Fitness Evaluation: For each candidate policy, the simulator runs multiple times ($T$ runs) to generate event logs. The fitness is calculated as the average cost vector across these runs to mitigate stochastic variance.
3. Genetic Operators: The NSGA-II algorithm applies selection, crossover, and mutation to evolve the population.
   • Selection: Uses non-dominated ranking and crowding distance to maintain diversity.
   • Crossover: Swaps entire probability distributions of decision points between parents to ensure valid offspring.
   • Mutation: Introduces novelty through four distinct variants:
     • Random: Reinitializes the probability distribution at a decision point.
     • Greedy: Aggressively shifts probability mass from the worst to the best successor based on a local heuristic.
     • Guided: Transfers a small, random fraction (10–50%) of probability mass from the worst to the best successor for fine-grained adjustments.
     • Hybrid: A weighted combination of the above strategies (60% Guided, 20% Greedy, 20% Random).

3. Output

The algorithm outputs a Pareto front of policies, offering practitioners a set of non-dominated solutions that represent different trade-offs between the defined objectives.

Key Contributions

• First Grounded Approach: This work introduces the first approach for grounded multi-objective optimization of resource-specific handover policies in business processes.
• Collaboration-Aware Optimization: Unlike prior work that treats resources as passive, this approach models resources as autonomous decision-making agents capable of choosing collaboration partners, thereby optimizing the collaborative structure of the process.
• Generic Framework: The approach is generic regarding the set of objectives, allowing for the discovery of Pareto fronts for any combination of performance metrics.
• Data-Driven Grounding: The simulation model can be discovered directly from event logs, ensuring the optimization is grounded in real process data.

Experimental Results

The approach was evaluated on synthetic and real-world datasets using labor cost and waiting time as the primary objectives.

Internal Validity (Synthetic Data)

Experiments were conducted on five variants of a Loan Application process with varying configurations (e.g., resource contention, skill differences, arrival rates).

• Performance: The proposed approach consistently outperformed the "as-is" baseline and heuristic baselines (Available, Random, Cheapest, Shortest Processing Time).
• Metrics: Across all scenarios, the approach reduced costs by an average of 37% and waiting time by 58% compared to the as-is process.
• Behavioral Analysis: In the LoanJS scenario, the optimized policy learned to balance the use of expensive, fast resources with cheaper, slower ones. In contrast, the "Cheapest" (LC) heuristic led to severe queue formation by exclusively using slow, cheap resources, while the optimized policy achieved a globally superior outcome through complementary collaboration patterns.

External Validity (Real-World Logs)

The approach was tested on seven widely used event logs from business process simulation research (e.g., P2P, ACR, BPI12W, BPI17W).

• Robustness: The approach improved both cost and waiting time compared to the as-is baseline in all seven datasets.
• Trade-offs: While the approach did not always achieve the absolute lowest cost (ranking above the LC baseline in four datasets), it consistently provided superior waiting times and a diverse set of Pareto-optimal solutions.
• Mutation Variants: The Guided mutation variant achieved the greatest average improvement (36.88% cost reduction, 57.71% waiting time reduction) across the datasets.
• Computational Cost: The optimization is computationally intensive, taking approximately 35 times longer than heuristic baselines (average 103.85 minutes vs. 2.96 minutes), though runtime scales similarly to the baselines with dataset size.

Significance and Claims

The paper claims that by shifting the optimization focus from global task assignment to resource-specific handover policies, organizations can leverage collaboration patterns to significantly improve process performance. The authors emphasize that their approach:

1. Captures Inter-Resource Dependencies: It explicitly models the decision-making autonomy of resources, a factor often ignored in standard RL or scheduling approaches.
2. Provides Actionable Alternatives: By generating a Pareto front, it offers decision-makers a transparent set of trade-offs rather than a single "optimal" solution, allowing them to select policies aligned with specific operational goals (e.g., prioritizing speed over cost).
3. Demonstrates Substantial Gains: The experimental results show that collaboration-aware optimization can yield substantial reductions in both cost and waiting time, outperforming established heuristics that optimize for single objectives or ignore long-term consequences.

The authors acknowledge limitations, including the reliance on simulation fidelity, the risk of overfitting to the simulator, and the computational overhead of the evolutionary search. They suggest future work could focus on robustness analysis, reducing computational costs, and extending the policy space to include dynamic parallelization.