NSGA-II optimization of resource-level handover policies cuts cost by 37% and waiting time by 58% on average across synthetic and real event logs

It presents the first multi-objective optimization of resource-level handover decisions, returning a set of person-specific handover policies on a Pareto front rather than a single global allocation rule. Existing RL and design-time methods treat resources as passive recipients of assignments, with decision variables being which resource executes which task; this work models resources as autonomous agents that choose collaboration partners, so the decision variable becomes the handover probability distribution π(a′|a,act,act′) between resources. The paper supports this distinction with formal definitions (handover policy as a conditional probability distribution, Pareto dominance) plus a method description, and contrasts it point by point against prior resource allocation, organizational mining, and agent-based simulation lines of work.

The method combines an existing MAS simulator (AgentSimulator) with NSGA-II and customizes crossover and four mutation operators to act on probability distributions, enabling search over probabilistic handover policies. Crossover swaps whole probability distributions at the level of entire decision points to keep distributions valid; mutation includes Random, Greedy, Guided, and Hybrid variants, where Guided transfers only 10%–50% of probability mass from the worst to the best successor and Hybrid mixes 60% guided, 20% greedy, and 20% random. The paper provides algorithm pseudocode (policy reconstruction, T simulation runs, averaging cost vectors) and hyperparameter settings (N=100, G=100, T=3, pc=70%, pm=30%), and compares the four mutation variants across seven logs.

Across five synthetic Loan Application variants the method improves both cost and waiting time over the as-is baseline in every scenario, and across seven real event logs it also improves cost and waiting time relative to as-is in all datasets. Relative to prior work, this is the first systematic reporting of improvements on the resource-level handover decision variable across both synthetic and real logs, compared simultaneously against four heuristic baselines (Avail, Random, LC, SPT). Internal validity uses five controlled variants (e.g., in LoanSCDT cost drops from $254k to $179k and waiting time from 2.78 to 0.60 hours); external validity uses seven public logs (e.g., ACR cost drops from $103k to $45k and waiting time from 8.97 to 0.51 hours), with Pareto-front quality metrics and runtimes also reported.

The method outputs a selectable set of trade-offs rather than a single point, for example a Pareto front on the real ACR log spanning roughly $22k–$85k in cost and about 0.2 to nearly 2 hours in waiting time. Compared with optimization methods that return one best solution, this gives process owners room to pick policies aligned with their operational goals, for instance a policy achieving about 0.2 hours when waiting time is prioritized. The paper uses the ACR Pareto-front figure and the 'most balanced solution' in the table ($45k, 0.51 hours) as reference points to describe front coverage and the magnitude of improvement over as-is, Avail, and LC.

The work targets design-time optimization of resource handover policies: the input is a multi-agent simulation model (specifiable manually with domain experts or discoverable automatically from an event log via AgentSimulator) plus any set of objective functions, and the output is a set of handover policies on a Pareto front. It applies to processes where resources have some decision-making autonomy over whom to collaborate with, particularly knowledge-intensive processes; the paper explicitly notes this premise may not hold for highly orchestrated workflow processes. Evaluation uses waiting time and labor cost as objectives and covers five synthetic Loan Application variants and seven public event logs spanning domains such as financial services, education, and procurement. For practitioners, what is directly usable is the candidate policies on the front and their trade-offs rather than a single 'optimal' answer; for researchers, the simulator is an interchangeable input artifact, so improvements in simulation fidelity can directly benefit optimization outcomes.

Optimization outcomes depend on simulation fidelity: the paper notes that if the simulator does not accurately reproduce real process dynamics, estimated improvements may not fully transfer to reality, and it suggests future work explicitly quantify simulator accuracy and analyze sensitivity of optimization results to modeling errors. Policies are optimized and evaluated within the same model, creating a risk of overfitting to simulator-specific dynamics, and the paper suggests robustness analysis across alternative simulation models or perturbed parameters. Real logs typically lack resource cost information, so the paper approximates it with performance profiles based on skills and activity durations, clustered by k-means into five tiers with hourly cost ranges (e.g., Tier 1: 10–25$/hr to Tier 5: 76–90$/hr), which may affect absolute performance estimates. On computation, baselines average about three minutes while the approach is roughly 35 times slower, with BPI17W taking 379.17 minutes in a single run. Evaluation also relies mainly on heuristic baselines, which the paper attributes to the lack of suitable existing benchmarks for the same handover optimization problem; this is a full-text read, but detailed definitions and results for the Pareto-front quality metrics are placed in the repository and only summarized in the paper.