Enterprise/PROGRESS.md

PROGRESS

2026-06-02 T-ALNS-RRD reproduction fidelity fix

Findings addressed

• The route travel component previously counted route elapsed time, which included waiting and service duration, instead of only sum t_ij(T_i) from the paper objective.
• RRD event generation and rollout used unseeded RNGs, so the same algorithm seed could produce different event streams and dispatch outcomes.
• Traffic incidents mutated the shared traffic tensor during detection, which could leak into later evaluations when an event was not dispatched.
• Rollout Tabu adjustment used the wrong sign for a cost-minimizing score: Tabu actions were rewarded and non-Tabu actions were penalized.
• Urgent-order actions could insert a negative placeholder node into fixed-size traffic tensors, causing invalid route evaluation.
• T-ALNS traffic aspiration did not compare against the current solution, and frequency memory did not use congestion-weighted updates from Eq.34.

Changes made

• Corrected Eq.1 cost accounting so travel cost is only arc traversal time; waiting and service remain part of time propagation.
• Documented hold-last-value traffic bucket behavior for customer windows after the paper's 6:00-18:00 traffic horizon.
• Restored congestion penalty generation to Eq.9 semantics: rho = theta * gamma.
• Threaded deterministic seeds through EventGenerator, RolloutEngine, and Dispatch.
• Made traffic incident severity temporary during dispatch evaluation and restored the base tensor with try/finally.
• Corrected rollout Tabu penalty/bonus direction for a cost-minimizing dispatch value.
• Reworked unsupported urgent-order insertion into fixed-graph penalty actions instead of invalid synthetic nodes.
• Added congestion-weighted frequency memory updates and current-solution traffic aspiration checks.
• Added configs/paper.yaml as the canonical paper-aligned server experiment config.

Local validation

• Per user instruction, no local Python tests, smoke tests, or full experiments were run in this round.
• Static review only: inspect source changes, config shape, and git diff before commit.

Server run command

cd t_alns_rrd_reproduction
pip install -r requirements.txt
python src/experiments/run_main_comparison.py --config paper --seeds 30 --iterations 1000 --time-limit 600 --output results/paper_fixed

Expected server outputs

• results/paper_fixed/tables/main_comparison.csv
• results/paper_fixed/tables/per_seed_costs.csv
• results/paper_fixed/tables/statistical_tests.csv
• results/paper_fixed/logs/convergence.npz

Interpretation rule

• Check whether Static > TA-Greedy > ALNS-Base > T-ALNS > T-ALNS-RRD in total cost, CES decreases, and OTDR improves.
• If the paper trend is not reproduced, keep the generated tables and record the failed metrics honestly instead of tuning results by hand.

2026-06-02 follow-up algorithm fidelity fix

Findings addressed

• ALNS, T-ALNS, and T-ALNS-RRD did not distinguish sigma_2 current-solution improvements from sigma_3 simulated-annealing accepted non-improving moves.
• T-ALNS diversification intensity used adaptive Tabu tenure instead of the paper's actual move-memory occupancy |T_move| / |T_move|max.

Changes made

• Reward assignment now follows the paper's four levels:
  • reward_global_best for new global best.
  • reward_improvement for accepted moves that improve the previous current solution.
  • reward_accepted for accepted non-improving moves.
  • reward_rejected for rejected moves.
• Added MoveTabu.utilization and changed Eq.27 diversification intensity to use actual move Tabu list occupancy.

Local validation

• Per user instruction, no local Python tests, smoke tests, or full experiments were run.
• Static checks only: source inspection and git diff --check.