Assortment Constraint Satisfaction

The solver guarantees that no solution violates non-negotiable business rules. These constraints are absolute and cannot be relaxed:

• Minimum Brand Representation: Ensures strategic partners always occupy a specified percentage of visible slots, preventing algorithmic bias from excluding key suppliers
• Shelf-Space Capacity: Respects physical or digital display limits, such as a maximum of 48 facings in a category page or 12 items in a recommendation carousel
• Regulatory Compliance: Enforces age-restricted product visibility, regional labeling requirements, or category-specific advertising bans

Unlike soft objectives that can be traded off, hard constraints trigger backtracking in the solver—if a proposed assortment violates a rule, the entire candidate solution is rejected and a new one is generated.

While constraints define the feasible solution space, the objective function drives the solver toward optimality. This mathematical expression quantifies what 'best' means:

• Revenue Maximization: Projects expected sales by multiplying predicted click-through rate, conversion probability, and item margin for each displayed product
• Profit-Aware Ranking: Incorporates supply chain costs, markdown liability, and holding costs into the scoring function
• Multi-Objective Optimization: Balances competing goals like revenue versus inventory turnover using Pareto frontier techniques, allowing merchandisers to explore trade-off curves

The solver iteratively evaluates thousands of candidate assortments, selecting the one that yields the highest objective score while satisfying all hard constraints.

Constraint propagation reduces the search space before the solver commits to a solution. This inference mechanism deduces the logical consequences of partial assignments:

• Forward Checking: When a product is tentatively placed in a display slot, the solver immediately eliminates all other products that would violate constraints if placed in adjacent slots
• Arc Consistency: Ensures that for every pair of constrained variables, every value in one variable's domain has a compatible value in the other's domain
• Domain Reduction: Shrinks the pool of eligible products for remaining slots as constraints tighten, dramatically accelerating convergence

This propagation prevents the solver from exploring branches of the search tree that are guaranteed to fail, reducing computation time from exponential to polynomial in many practical retail scenarios.

When the solver reaches a dead end where no valid product can fill a remaining slot, it employs intelligent backtracking rather than naive chronological backtracking:

• Conflict Analysis: Identifies the specific combination of earlier assignments that caused the failure, not just the most recent decision
• Conflict Clause Generation: Records the conflicting pattern as a new constraint, preventing the solver from repeating the same mistake in different branches of the search tree
• Backjumping: Jumps directly to the most recent decision variable involved in the conflict, skipping over irrelevant intermediate choices

This technique, borrowed from SAT solver research, transforms the solver from a brute-force enumerator into a learning system that becomes more efficient as it explores the solution space.

The order in which the solver assigns products to display slots dramatically impacts performance. Sophisticated heuristics guide this process:

• Most Constrained Variable First: Prioritizes slots with the fewest eligible products remaining, such as a 'Featured Brand' position with strict contractual requirements
• Most Promising Value First: Within a slot, tries products with the highest expected revenue contribution before lower-value alternatives, increasing the chance that the first complete solution is near-optimal
• Fail-First Principle: Deliberately tackles the hardest subproblems early, triggering conflict learning when the search tree is still shallow and the cost of backtracking is low

These heuristics are often learned from historical solver traces using reinforcement learning, adapting to the specific constraint patterns of each retailer's catalog.

Rather than solving each assortment problem from scratch, the solver reuses structure from prior optimizations:

• Solution Caching: Stores the optimal assortment for common constraint configurations, such as a specific store cluster during a weekday afternoon
• Incremental Re-Optimization: When a single constraint changes—like a product selling out—the solver adjusts the existing solution rather than rebuilding it entirely
• Similarity-Based Initialization: Seeds the search with the solution from the most similar historical context, using k-nearest neighbors over constraint feature vectors

Warm starting reduces median solve time by 60-80% in production environments, enabling real-time assortment updates within the sub-100ms latency budgets required for online serving.

A luxury fashion platform must guarantee that Nike and Adidas each occupy at least 15% of the visible slots on a sneaker category page, regardless of their real-time click-through rates. The constraint solver treats this as a hard cardinality constraint, ensuring the final assortment never violates the contractual agreement. If Nike inventory drops below the threshold, the solver triggers a stockout probability check and requests a substitute from the same brand tier before removing the slot.

A grocery chain's mobile app displays a 'Fresh Produce' carousel limited to exactly 12 items due to UI constraints and cognitive load research. The solver must select the optimal 12 products from 200 available SKUs while respecting:

• Perishability windows: Items expiring within 24 hours get priority boosting
• Local sourcing rules: At least 4 slots must feature regional farms
• Profit margin floors: No item below 22% margin can occupy a slot The solver frames this as a bounded knapsack problem with a hard capacity of 12.

A home improvement retailer's search results page must never show more than 3 items from the same subcategory in the top 10 results for a generic query like 'lighting.' This prevents a single brand or product type from dominating the viewport. The constraint solver uses a diversity constraint that iterates through the ranked list, swapping out the 4th occurrence of 'floor lamps' with the next-best 'pendant light' or 'smart bulb' candidate, maintaining relevance while enforcing variety.

A global alcohol delivery platform must suppress all CBD-infused beverages from the assortment when a user's GPS coordinates fall within Utah state boundaries. The solver applies a geofenced exclusion rule as a pre-filtering step before any ranking model runs. This hard constraint operates at the query level, removing entire SKU groups from the candidate pool. The system logs the suppression event for audit trails required by state regulators.

An electronics retailer's 'Compare Similar Items' widget must display exactly 4 alternative laptops when a shopper views a specific MacBook model. The constraint solver enforces:

• Price band: Alternatives must be within ±15% of the seed product's price
• Stock availability: All 4 alternatives must have >5 units in the local warehouse
• Brand exclusion: Cannot show another Apple product to avoid cannibalization If fewer than 4 candidates satisfy all constraints, the solver widens the price band to ±25% rather than showing empty slots.

During a Black Friday event, a marketplace must allocate 3 premium banner slots on the homepage. The solver balances multiple hard constraints:

• Paid placement contracts: Slot 1 is reserved for the highest-bidding brand partner
• Inventory depth: No promoted product can have fewer than 1,000 units in the distribution center
• Profit protection: The combined margin of all 3 promoted items must exceed 40% The solver treats this as a constraint optimization problem, maximizing expected revenue while treating the contractual and inventory rules as non-negotiable.