A shielding mechanism lets deep RL learn discriminatory pricing while provably enforcing per-period price-fairness bounds; in simulated markets Shield-SAC matches high revenue while never violating fairness constraints.

Main Finding

Introducing an optimization-based shielding mechanism into a model-free deep reinforcement learning (DRL) pricing agent yields a practical algorithm—Shield Soft Actor-Critic (Shield-SAC)—that enforces instantaneous, hard inter-group price-fairness constraints at every period while achieving strong revenue performance in simulated perishable-inventory markets.

Key Points

• Problem: Dynamic pricing for perishable products with limited inventory over a finite horizon and heterogeneous customer groups. Firms may want discriminatory prices but must bound inter-group price differences to avoid legal risk and consumer backlash.
• Formalization: The pricing problem with per-period fairness constraints is cast as an action-constrained Markov decision process (ACMDP) with unknown demand functions.
• Challenge: Off-the-shelf model-free DRL algorithms (for unconstrained MDPs) do not guarantee that executed actions satisfy hard, instantaneous constraints.
• Solution (Shielding): An optimization-based shield intercepts the agent’s proposed price vector and enforces the per-period fairness constraints before execution. From the agent’s viewpoint this induces a shield-induced unconstrained MDP, but in reality all executed prices satisfy the constraints.
• Algorithm: Shield-SAC—combines Soft Actor-Critic (a model-free DRL method) with the optimization-based shield to produce safe pricing policies.
• Empirical result: In two simulated markets of different scale, Shield-SAC consistently enforces price-fairness constraints both during training and deployment while maintaining strong revenue performance relative to unconstrained baselines.

Data & Methods

• Environment: Simulated finite-horizon selling problems with perishable inventory and heterogeneous customer groups; demand functions treated as unknown to the learning agent.
• Modeling: ACMDP formulation where feasible action sets at each time step are constrained by hard, instantaneous bounds on inter-group price differences (price fairness constraints).
• Learning framework: Model-free DRL (Soft Actor-Critic) trained to maximize revenue; agent proposals are filtered by an optimization-based shield that guarantees feasibility.
• Shielding mechanism: An online optimization layer that ensures executed actions satisfy the fairness constraints at each period (converts candidate actions into feasible ones). This allows safe interaction with the environment throughout training and deployment.
• Evaluation: Experiments on two simulated market instances (different scales) comparing revenue and constraint violations; metrics reported include revenue achieved and whether fairness constraints were ever violated.

Implications for AI Economics

• Practical safe RL for pricing: Demonstrates a tractable pathway to deploy data-driven dynamic pricing while respecting legally and socially motivated fairness constraints—critical for real-world adoption in airlines, hospitality, retail, etc.
• Revenue–fairness trade-offs: Shows it is possible to uphold hard fairness limits without catastrophic loss in revenue, narrowing the practical trade-off and making constrained pricing more viable.
• Regulatory compliance & consumer trust: Constraint-guaranteeing mechanisms like shielding can help firms comply with regulation and avoid backlash, enabling more widespread use of automated pricing systems.
• Methodological transfer: The shielded-DRL approach can generalize to other constrained decision problems in economics (capacity caps, service-level constraints, demand-side fairness) where hard instantaneous constraints must be enforced.
• Future directions: Validate on field data and live A/B tests, explore other fairness definitions (longitudinal or distributional fairness), analyze computational cost and scalability of the shield optimization, and study robustness when demand nonstationarity or model mismatch occurs.