prasanthdj8/retail-inventory-openenv

Retail Inventory & Expiry Management — OpenEnv

An OpenEnv environment that simulates a store manager making real-time inventory decisions for perishable products. The AI agent must decide when to discount, reorder, or remove items nearing expiry, balancing sales revenue, food waste reduction, and stockout avoidance.

Why this environment?

Retailers, supermarkets, and pharmacies lose billions annually to expired stock and inefficient ordering. This environment models the daily decision-making challenge of a store manager — a genuine, high-impact task for which RL agents can provide real value.

Environment overview

Action space

Observation space

Each observation contains:

• —day / total_days — current day and episode length
• —products — list of product snapshots:
• —product_id, name, category
• —stock — units currently on shelf
• —price, cost — current price and unit cost
• —days_to_expiry — days until expiry (-1 = non-perishable)
• —current_discount — active discount percentage
• —demand_estimate — deterministic daily demand estimate
• —is_expired — whether product has already expired
• —daily_revenue, daily_waste_cost — today's financials
• —cumulative_revenue, cumulative_waste_cost — episode totals
• —stockout_events — products that ran out today
• —budget_remaining — reorder budget left

Reward function

reward = sales_revenue − waste_penalty − stockout_penalty − reorder_cost

• —`sales_revenue`: units sold × effective price (positive signal every day)
• —`waste_penalty`: −cost × expired units (penalises letting items expire)
• —`stockout_penalty`: −5.0 per stockout event (penalises running out)
• —`reorder_cost`: −cost × reorder_qty (immediate reorder expenditure)

The reward is dense — partial progress is rewarded every day, not just at episode end.

Tasks

Easy — Single Product Management

• —Products: 1 perishable product
• —Horizon: 7 days
• —Goal: Sell as much as possible before expiry; avoid waste
• —Grader: score = revenue_ratio × (1 − waste_ratio) ∈ [0, 1]

Medium — Multi-Product Balance

• —Products: 5 perishable products
• —Horizon: 30 days
• —Goal: Balance revenue, waste reduction, and stockout avoidance
• —Grader: weighted combination (revenue 40%, waste 35%, stockouts 25%)

Hard — Full Store Optimisation

• —Products: 20 perishable products across multiple categories
• —Horizon: 90 days
• —Goal: Maximise profit margin + sustainability bonus for near-zero waste
• —Grader: weighted combination (profit 35%, waste 30%, stockouts 20%, sustainability 15%)

All graders return deterministic, reproducible scores in [0.0, 1.0].

API endpoints

Setup & Usage

pip install -r requirements.txt
python main.py

docker build -t retail-inventory-openenv .
docker run -p 7860:7860 retail-inventory-openenv

Baseline inference

export HF_TOKEN=your_token
export MODEL_NAME=meta-llama/Llama-3.1-8B-Instruct
export API_BASE_URL=https://router.huggingface.co/v1
python inference.py

License

MIT