Curious Now

Story

Surrogate-Assisted Genetic Programming with Rank-Based Phenotypic Characterisation for Dynamic Multi-Mode Project Scheduling

ComputingMaterials & Engineering

Key takeaway

Researchers developed a new algorithm to help manage changing project schedules and resource constraints more efficiently. This could improve how organizations plan and adapt large, complex projects.

Read the paper

Quick Explainer

The researchers proposed a new approach to enable surrogate-assisted genetic programming for dynamic, resource-constrained project scheduling. The key idea is to capture the ranking behavior of the scheduling heuristics, rather than just their fitness values, and use this as a surrogate model to estimate the quality of new heuristics generated by the genetic programming process. This rank-based phenotypic characterization allows the surrogate model to effectively guide the search towards high-performing scheduling policies, while requiring fewer expensive fitness evaluations compared to prior methods. The distinctive aspect is the suitability of this approach for the dynamic, multi-mode project scheduling problem, which involves both ranking and subset selection of eligible activities.

Deep Dive

Technical Deep Dive: Surrogate-Assisted Genetic Programming for Dynamic Multi-Mode Project Scheduling

Overview

This research proposes a new phenotypic characterization (PC) scheme to enable surrogate-assisted genetic programming (GP) for the dynamic multi-mode resource-constrained project scheduling problem (DMRCPSP). The PC scheme captures the ranking behavior of GP heuristics over eligible activities and activity groups, allowing a surrogate model to estimate the fitness of unevaluated GP individuals.

Problem & Context

  • DMRCPSP requires making real-time scheduling decisions under changing project states and resource availability
  • GP can effectively evolve heuristic rules for DMRCPSP, but the evolutionary process relies on many expensive fitness evaluations via simulation
  • Surrogate models can reduce evaluation cost, but require problem-specific PC schemes to capture heuristic behavior
  • Existing PC schemes are not suitable for DMRCPSP due to its unique decision structures involving both ranking and subset selection

Methodology

  • Proposed a rank-based PC scheme that extracts the ranking behavior of GP heuristics across representative decision situations
  • Integrated the PC scheme into a surrogate-assisted GP algorithm, using nearest-neighbor regression to estimate fitness of intermediate offspring
  • Introduced an offspring multiplier parameter to control the number of intermediate offspring generated per generation

Data & Experimental Setup

  • Evaluated on 3 DMRCPSP scenarios with varying precedence complexity and resource constraints
  • Compared the proposed SKGGP algorithm against the baseline KGGP algorithm from prior work
  • Ran 30 independent trials of each algorithm with 100 generations and 1000 population size

Results

  • SKGGP with larger offspring multipliers (2, 4) outperformed KGGP across all scenarios
  • SKGGP could identify high-quality heuristics using 20-40% fewer expensive fitness evaluations
  • Surrogate precision decreased as offspring multiplier increased, but still enabled effective offspring selection

Limitations & Uncertainties

  • Surrogate model's precision limited when selecting from a large number of intermediate offspring
  • Expanding the surrogate database and exploring more expressive PC schemes could improve accuracy

What Comes Next

  • Investigate techniques to maintain surrogate precision with larger offspring populations
  • Explore alternative machine learning models for the surrogate beyond nearest-neighbor regression
  • Assess the proposed approach on additional DMRCPSP problem instances and configurations

Source