Video summary

Webinar Bidang Kajian Supply Chain Modeling and Analytics ISLI

Main summary

Key takeaways

Educational

Main ideas and lessons conveyed

Overall webinar theme

The webinar discusses Supply Chain Modeling and Analytics (ISLI) topics, with emphasis on:

  • Logistics transportation/distribution optimization
  • Network expansion and hub location
  • Machine learning + optimization frameworks for scheduling
  • Structural Equation Modeling (SEM) / PLS-SEM for modeling perceptions and latent constructs

A recurring theme is choosing the right modeling/solution approach—for example:

  • heuristic vs metaheuristic vs exact methods vs ML/SEM based on problem complexity, data availability, and objectives.

Speaker 1: Mr. David Triliputra (Heuristic & Metaheuristic for VRP)

Focus/problem addressed

A Vehicle Routing Problem (VRP) variant:

  • Asymmetric Capacitated VRP with Time Windows
  • Limited vehicle capacity
  • Asymmetric distances
  • Customer time limits (hard time windows)
  • Regular vs overtime working hours:
    • Regular hours = soft time window
    • Overtime limit = hard constraint
  • Operational issues addressed:
    • Delivery delays
    • Overtime usage

Goal: minimize total shipping cost.

Model components

Objective function

Minimize total shipping cost, composed of 4 cost components:

  1. Travel costs

    • Travel distance × fuel cost × fuel consumption.

  2. Overtime shipping cost

    • Overtime duration × overtime cost rate.
    • Overtime occurs when delivery time exceeds regular working hours.
    • Delivery time includes:
      • travel time
      • service/unloading time
      • driver break time

  3. Late compensation fee

    • Applied when deliveries exceed customer deadlines.
    • Based on: “late quantity” × compensation cost (approximated using a portion of unit product price).

  4. Return shipping cost

    • When a customer rejects delivery due to lateness, goods must be returned.
    • Return logistics increase effective distance/cost.

Constraints (as described)

  • Vehicle capacity not exceeded.
  • Set partitioning structure:
    • each vehicle serves only one route
    • each reverse route served by only one vehicle
  • Do not exceed:
    • customer delivery deadlines (hard time window)
    • overtime limits
  • Binary decision variable constraints.

Solution methodologies (detailed algorithm steps)

A) Heuristic method: Sequential Insertion (SI)

Classification: constructive heuristic (builds routes from the start).

Core idea

  • Sequential insertion: typically selects closest customers in each step (nearest/greedy-like behavior).
  • The overall method iterates rather than using only local improvements.

General steps described

  1. Initialize:
    • create initial route(s)
    • mark all customers as not visited
  2. While unserved customers remain:
    • select the closest customer to the warehouse
    • compute:
      • total demand if inserted
      • total delivery time (including travel/service/break structure)
    • check feasibility:
      • capacity satisfied?
      • customer time windows satisfied?
      • overtime time window satisfied?
    • if feasible:
      • insert customer into the route
      • mark visited
  3. Start a new route and repeat until all customers are served.

Real-world enhancement: vehicle delivery-time balancing

  • Two vehicles’ total delivery times should be balanced.
  • Let TA be allowed tolerance (company tolerance ≈ ±30 minutes).
  • Procedure:
    • compute delivery time difference between vehicles
    • if difference > TA:
      • move the last customer from the route with longer delivery time to the route with shorter delivery time
      • re-check constraints and update cost/time
    • repeat until balanced within TA (or no more feasible balancing)

B) Metaheuristic method: Ant Colony Optimization (ACO)

A comparison metaheuristic.

Key ACO parameters mentioned

  • number of ants (amax)
  • pheromone trail values (τᵢⱼ)
  • evaporation rate
  • alpha (pheromone importance)
  • beta (visibility/heuristic importance)

Intuition

  • Ants construct routes probabilistically.
  • Pheromones guide future ants toward promising routes.
  • Evaporation prevents premature convergence and encourages exploration.

General ACO process described

  1. Initialize:
    • ACO parameters (ants, evaporation, alpha, beta)
    • compute visibility values
  2. For each iteration:
    • build route(s) probabilistically:
      • start at the warehouse
      • compute selection probability for each next customer:
        • based on pheromone + distance/time influence
      • use random sampling with cumulative probabilities to choose next customer
    • feasibility check:
      • if constraints satisfied (capacity, time windows, overtime), accept; otherwise reject/cancel that insertion
    • create next route if needed
  3. After construction:
    • update pheromones based on route quality (distance/cost)
    • include time-balancing logic iteratively (similar to SI balancing approach)

Case study results summarized

Context: clothing distribution/convection industry in Bandung.

Data

  • 16 shops (customers)
  • 2 vehicles
  • vehicle capacity: 2000 units
  • regular time limit: 480 minutes
  • overtime limit: 540 minutes
  • customer time limit: 510 minutes

Findings

  • SI significantly reduced costs vs the baseline:
    • first route cost decreased by eliminating overtime and delays
    • second route also improved
  • ACO improved further:
    • total cost reductions slightly more than SI
  • Delivery time balancing comparison
    • SI achieved very balanced vehicle times (e.g., 508 vs 510 minutes)
    • ACO improved cost more but showed a larger time difference in one route
  • Conclusion: ACO performs better overall for cost and global solution quality, especially as problem size grows (hundreds/thousands customers).

Takeaways from Speaker 1

  • Heuristics (SI)
    • simpler
    • faster
    • good when problems are relatively small/simple
  • Metaheuristics (ACO/EO-like family)
    • better at escaping local optima
    • better for more complex combinatorial optimization
  • Time balancing constraints can materially affect operational usefulness.

Speaker 2: Mr. Fran Setiawan (Hub Network Expansion via Benders Decomposition)

Problem domain

  • Hub Location / Hub Network Design / Hub Network Expansion
  • Focus on Hub Network Expansion:
    • start with an existing hub network and decide:
      • which additional hubs to add
      • which additional links to build
      • where economies of scale apply

Conceptual model

Classical hub location:

  • hubs act as consolidation points
  • spokes allocated to hubs
  • transportation between hubs benefits from economies of scale

Economies of scale modeled via alpha (0 to 1):

  • affects transportation costs between hubs.

Extension: Hub Network Design

Removes classical assumptions:

  • hubs do not need to be fully interconnected
  • direct links between spokes may be allowed
  • building hubs/links incurs fixed costs

This forms the basis for “expansion.”

Hub Network Expansion (what gets decided)

Given:

  • existing hubs
  • existing links

Decide:

  • add new hub nodes and new links
  • route commodities through the resulting hub network

Costs included:

  • fixed cost to build a hub
  • fixed cost to build a link
  • routing/transportation costs depending on whether movement is hub↔hub or not

Mathematical structure (high-level)

Sets

  • N nodes
  • hub set J’ (existing hubs)
  • link set E (all edges) and E’ (existing edges)
  • commodities P with OD demands

Variables

  • binary variables for building hubs and links
  • continuous flow/proportion variables for commodity routing on arcs

Objective

  • minimize total cost = hub build + link build + routing cost (with scale factor alpha)

Constraints (described)

  • existing hubs must remain selected
  • flow conservation and commodity routing from origin to destination
  • if a commodity uses a link/arc, corresponding hub/link must exist (activation constraints)

Expansion is NP-hard, making exact solution difficult for large cases.

Solution approach: Benders Decomposition (exact method)

Core idea

Decompose into:

  • Master problem: integer/binary decisions (hub/link selections)
  • Subproblem: route commodities given fixed master decisions (continuous)

Iteration

  1. Solve master to determine Y/Z hub/link selections.
  2. Solve subproblem(s).
  3. Derive Benders cuts.
  4. Add cuts to master and repeat until convergence.

Enhancements described to improve performance

  1. Multicut
    • decomposes subproblems per commodity for multiple cuts.
  2. Two-phase approach
    • phase 1: relax integrality (solve LP relaxation for y/z)
    • phase 2: restore integrality and iterate for mixed-integer optimality
  3. Pareto-optimality cuts
    • stronger cuts to speed convergence (reduce iterations)
  4. Combination variants (as stated)
    • naming like BMT / BMTP represents combinations such as:
      • multicut + two-phase
      • pareto optimality cut
      • combined strategy (multicut + two-phase + pareto cuts)

Experimental findings (summary)

  • Compared Benders variants against CPLEX.
  • Reported emphasis:
    • on some hard instances, Benders variants outperform or solve instances CPLEX can’t within time limits
    • high hub fixed costs increase computational difficulty

Case study: Indonesia sea toll road / port expansion

  • 2015 initiative: sea toll road concept
  • plan: ~24 strategic ports
  • existing operational ports listed:
    • Belawan, Tanjung Perak, Tanjung Priok, Makassar
  • model uses province-level nodes (up to 38 provinces) and scenario-based alpha values

Reported insights

  • lower alpha → stronger economies of scale → more hub/link additions
  • cost reduction via economies of scale can be substantial (up to ~49.8% in scenario)
  • hub expansion reduces congestion flows into major Java ports (large reduction percentages reported)

Takeaways from Speaker 2

  • Hub expansion is highly sensitive to:
    • alpha (economies of scale)
    • fixed vs routing cost sensitivity
  • Benders decomposition is appropriate because routing becomes easier once hub/link selection is fixed.
  • Managerial uses:
    • scenario analysis (alpha sensitivity)
    • deciding when hub expansion is worthwhile beyond adding only links.

Speaker 3: Mr. Adityya Saputra (ML for “Problem Importance Analytics” in 3-stage Export Container Scheduling)

Domain/problem

Three-stage export container scheduling:

  • yard/container gathering (yarding)
  • internal transport (internal truck)
  • quay/crane stage (quay crane)

Objective: minimize maximum completion time (maxpan / latest completion).

Motivation/lesson

  • Traditional optimization treats subproblems uniformly.
  • This work asks:
    • which subproblem (yard vs internal transport vs crane) most influences overall output?
  • Then iteratively uses ML to decide which subproblem to solve next.

Proposed framework (explicit multi-part workflow)

Part 1: Identify high-impact subproblems

  • Estimate each subproblem’s effect on overall output.
  • Use ML to identify which input variables most predict maxpan.

Part 2A: Define optimization problems

  • Define a scheduling/optimization subproblem for each stage.

Part 2B: Problem selection model

  • Use the ML-informed model to choose which subproblem(s) to solve.

Part 2C: Solve chosen subproblem

  • Solve selected optimization subproblem(s) under operational constraints.

Re-evaluation loop

  • After solving and observing the resulting system outcome:
    • feed outcome back to the selection model
    • pick a new subproblem set again (iterative improvement)

Scheduling setup described

  • Containers scheduled across stages with constraints:
    • each stage processes one container at a time
  • Flow precedence constraints:
    • internal transport cannot start until yarding completes for that container
    • crane cannot serve until internal transport completes
  • Scheduling rules used in the example:
    • FCFS (first-come-first-served)
    • SPT (shortest processing time)

Machine learning input/output structure

Inputs

  • summary statistics and indicators, such as:
    • averages and standard deviations for processing times per stage
    • ratios between resources (yard/truck, truck/crane)
    • selected scheduling rule types per stage

Output

  • label: max completion time / maxpan

Implementation (Python / Google Colab)

Steps mentioned:

  1. import libraries (pandas noted)
  2. read dataset from Excel into a dataframe
  3. visualize correlation:
    • Spearman correlation heatmap
  4. data preparation:
    • clean empty values
    • split features vs target
    • train/test split (80/20)
    • standard scaler normalization
  5. train regression models:
    • linear regression
    • support vector regression
    • random forest
    • KNN
    • decision tree
    • gradient boosting
  6. evaluate with metrics:
    • RMSE, MAE, MSE, R²
  7. select best model based on threshold criteria (including an idea around 70% of R² as discussed) and ranking

Results: “problem importance” findings

  • Correlation heatmap highlights key variables linked to maxpan, especially:
    • crane-related rule
    • truck-related rule/time
    • yarding rule appeared least influential
  • Feature importance (using the chosen top model—linear regression):
    • highest influence crane-related factor (reported ~25.12)
    • truck rule next (~2.22)
    • yarding rule lowest

Takeaways from Speaker 3

  • ML can support decision-making by identifying which stage/subproblem most impacts overall performance.
  • This reduces the risk of “solving the wrong subproblem first” and improves efficiency of optimization iterations.
  • Positioned as a proof-of-concept likely generalizable to other logistics/scheduling contexts by:
    • defining an output metric
    • defining features related to sub-stages
    • learning which sub-components dominate the output.

Speaker 4: Mrs. Resista Vika Liana (SEM / PLS-SEM Concepts and Applications)

Topic

Structural Equation Modeling (SEM), with emphasis on PLS-SEM.

SEM models relationships between:

  • observed variables (indicators/items)
  • latent constructs (unobserved variables)

Strong emphasis: SEM is widely used for perception-based research in logistics and beyond.

Conceptual progression described

  • starts from:
    • regression (X → Y) limitations
  • then:
    • path analysis / path modeling
  • then:
    • SEM to handle latent variables and more complex structures
  • SEM described as:
    • a combination of path modeling + regression + latent constructs
  • “first generation vs second generation” SEM terminology mentioned.

Types of SEM emphasized

  • CB-SEM vs PLS-SEM
    • CB-SEM: covariance-based, theory confirmation; needs stronger assumptions/theory validation
    • PLS-SEM: variance-based; flexible for prediction/exploration; can extend/modify theory without fully rejecting it

Advantages of PLS-SEM stated

  • works with smaller sample considerations (noted later in standard rules)
  • nonparametric stance (fewer distribution assumptions)
  • can use primary survey/measurement data
  • reusability:
    • adjust one structural model for other studies with modifications

SEM measurement concepts

  • latent constructs measured through indicators/items
  • exogenous vs endogenous variables
  • reflective vs formative indicator logic:
    • reflective: latent variable causes indicators
    • formative: indicators form/build the latent construct

Tooling

  • SmartPLS
  • alternatives mentioned:
    • JASP / R packages (coding-based open-source approach)

SEM examples mentioned briefly

  • supply chain performance model extended with transportation factors
  • consumer behavior model combining multiple theories
  • halal awareness model with mediating/moderating roles
  • general emphasis: SEM can integrate multiple theories by adjusting the structural model.

Takeaways from Speaker 4

  • SEM/PLS-SEM provides a structured way to model latent constructs and relationships, useful for logistics studies involving:
    • perception
    • acceptance
    • behavioral constructs
  • Conceptually, it can be combined with other methods; practically, SEM often serves as the modeling layer for latent relationships, while other methods address other quantitative optimization tasks.

Methodologies and instruction-like content (consolidated)

A) Heuristic SI for VRP (route construction + balancing)

  • Initialize routes and mark customers unvisited.
  • Repeat until all customers are served:
    • select the closest candidate customer to the warehouse (greedy insertion choice)
    • compute demand impact and delivery time impact (including service/unloading + driver break)
    • if constraints satisfied (capacity, hard time windows, overtime limits), insert customer
    • mark customer visited
    • when route cannot accept more, start a new route
  • Apply time balancing:
    • compute difference between total delivery times of vehicles
    • if difference > TA (tolerance ≈ ±30 minutes):
      • move last customer from longer route to shorter route
      • re-check constraints/costs
    • repeat until time difference is within TA

B) ACO metaheuristic for VRP

  • Initialize pheromone levels, evaporation rate, alpha, beta.

Original video