Description

This notebook is designed to turn the VN2 simulator into an optimization loop for actionable ordering policies.

What it does

Simulator wrapper: Weekly transition aligned with the organizer’s logic (no backorders, lead time 2, weekly review, holding on EndInv, shortage on missed sales).

Three submission paths:

• Baseline parametric base‑stock (grid‑tuned k and service level) → orders_round1_rl.csv.
• Segment‑wise tuning (Department-level k and service level via Optuna) → orders_round1_rl_segments.csv.
• Feature‑driven direct policy (k learned from recent sales and low‑inventory features via Optuna) → orders_round1_rl_policy.csv.

Key assumptions

• StartInv = EndInv_prev + InTransit_W+1_prev; new W+1 ← prev W+2; new W+2 ← Orders.
• Protection period P = 3 (lead 2 + review 1).
• Costs: 0.2 per unit of EndInv (holding), 1.0 per unit of missed sales (shortage).

Inputs (Week 0)

• Sales: “Week 0 - 2024-04-08 - Sales.csv”
• Initial State: “Week 0 - 2024-04-08 - Initial State.csv”
• Master (for segments): “Week 0 - Master.csv”
• Submission Template: “Week 0 - Submission Template.csv”

How to run

• Run top‑to‑bottom:

1) Setup and simulator instantiation.2) Baseline stats/policy + grid evaluation; emit orders_round1_rl.csv.3) Optuna segment‑wise tuning; emit orders_round1_rl_segments.csv.4) Optuna feature policy search; emit orders_round1_rl_policy.csv.

• Notes are embedded before submission cells to indicate dependencies (e.g., finishing the corresponding Optuna study).

Tuning tips

• Increase Optuna n_trials for better results; adjust segment key (e.g., ProductGroup) by changing the Master join.
• Extend the feature policy with richer signals (lags, rolling CV, availability) or switch to a small neural network for k.
• Keep TBATS/complex models for top seasonal SKUs to manage runtime.

Repository

• Full project and CLI: https://github.com/senoni-research/vn2inventory