Chance Constraints

A chance constraint does not demand absolute, deterministic satisfaction of a constraint. Instead, it requires the constraint to hold with at least a user-specified probability or confidence level (e.g., 95%). Formally, for a constraint $g(x) \leq 0$, a chance constraint is written as $\mathbb{P}(g(x) \leq 0) \geq 1 - \epsilon$, where $\epsilon \in (0,1)$ is the acceptable risk of violation. This directly trades off performance (allowing more aggressive control) with safety risk.

Chance constraints are explicitly designed for systems with stochastic uncertainty. This uncertainty can originate from:

• Additive noise on measurements or dynamics.
• Parametric uncertainty in the model (e.g., uncertain friction coefficient).
• Forecast errors for external disturbances (e.g., wind gusts). Unlike worst-case Robust MPC, which must plan for the absolute worst possible realization of uncertainty (leading to conservative, potentially infeasible controllers), chance constraints allow for rare, low-probability constraint violations, resulting in more performant control policies.

The probabilistic constraint $\mathbb{P}(g(x, \xi) \leq 0) \geq 1 - \epsilon$ is typically intractable to enforce directly within an online optimization. The core engineering challenge is to reformulate it into an equivalent or approximating deterministic constraint that can be handled by standard solvers. Common reformulation methods include:

• Boole's inequality (Union Bound) for multiple constraints.
• Assuming a specific distribution (e.g., Gaussian) and using the inverse cumulative distribution function.
• Using distributionally robust approaches that only require knowledge of moments (mean, variance).

A standard result for linear constraints with Gaussian uncertainty is that a chance constraint can be enforced by tightening the original deterministic constraint. For a constraint $a^T x \leq b$ where $x$ is uncertain with covariance $\Sigma$, the tightened deterministic constraint becomes $a^T \mathbb{E}[x] \leq b - \Phi^{-1}(1-\epsilon) \sqrt{a^T \Sigma a}$, where $\Phi^{-1}$ is the inverse Gaussian CDF. This creates a smaller, safer feasible region within the optimizer, where the back-off amount scales with the uncertainty magnitude ($\sqrt{a^T \Sigma a}$) and the desired confidence level.

Chance constraints are a specific, simple form of a risk measure. The probability of violation $\epsilon$ is the risk level. More advanced risk measures used in stochastic control and finance, such as Conditional Value-at-Risk (CVaR), can be integrated into MPC frameworks. CVaR considers the expected severity of violations when they occur, providing a more coherent measure of tail risk than a simple probability. This makes the controller attentive not just to the frequency of violations, but also to their potential magnitude*.

Chance-constrained MPC is critical in domains where uncertainty is inherent and absolute safety is either impossible or too costly to guarantee.

• Autonomous Vehicles: Ensuring a collision probability remains below an acceptable threshold despite sensor noise and prediction errors for other agents.
• Energy Management: Managing battery state-of-charge constraints with uncertain renewable energy generation and load forecasts.
• Process Control: Maintaining safe temperature or pressure limits in chemical reactors with uncertain reaction kinetics or feedstock quality.
• Robotics: Executing manipulation tasks near obstacles despite state estimation uncertainty and model inaccuracies.

Chance constraints are critical for safe navigation in stochastic environments. They allow a vehicle's MPC to plan trajectories that guarantee a collision probability below a strict threshold (e.g., 99.9%) despite uncertain pedestrian motion or sensor noise.

• Key Mechanism: The constraint on vehicle position relative to an obstacle is formulated as Pr(position ∈ Safe Set) ≥ 0.999.
• Benefit: Enables less conservative, more efficient maneuvering than worst-case robust MPC, which would assume all obstacles move with maximum adversarial speed.

In renewable energy systems, chance constraints manage the uncertainty of solar/wind generation and load demand. They ensure battery state-of-charge limits and power flow constraints are met with high probability over a receding horizon.

• Example Constraint: Pr(SoC(t) ≤ SoC_max) ≥ 0.95 ensures the battery is not overcharged with 95% confidence, given probabilistic forecasts.
• Outcome: Allows for more aggressive use of variable renewables, increasing economic efficiency while maintaining grid reliability.

For robots grasping objects with unknown mass or friction properties, chance constraints in MPC ensure feasible actuator torques and stable contact forces. This is vital for assembly and logistics tasks.

• Application: A constraint on joint torque: Pr(τ(t) ≤ τ_max) ≥ 0.99 accommodates uncertainty in the object's inertial parameters.
• Implementation: Often uses affine disturbance feedback parameterization to reformulate the probabilistic constraint into a deterministic, tractable form for real-time optimization.

Chemical reactors have strict safety bounds on temperature and pressure. Chance constraints enable optimal economic operation despite uncertainties in reaction kinetics and feedstock quality.

• Use Case: Maintaining reactor temperature below a critical threshold: Pr(T(t) ≤ T_critical) ≥ 0.999.
• Method: The uncertainty distribution is often approximated using scenario-based approaches or distributionally robust optimization, where the constraint must hold for a family of possible distributions.

A probabilistic constraint Pr(g(x, ξ) ≤ 0) ≥ 1-ε is not directly solvable. Key reformulation methods include:

• Assumption-Based Reformulation: If uncertainty ξ is Gaussian, the constraint can become a deterministic second-order cone constraint.
• Scenario Approach: Replace the probabilistic constraint with N hard constraints for N random samples. The sample count N is chosen via probably approximately correct (PAC) bounds to guarantee the original chance constraint.
• CVaR Approximation: Use the Conditional Value-at-Risk as a convex, conservative approximation of the chance constraint, leading to a tractable convex program.

Chance constraints provide a less conservative alternative to Robust MPC.

• Robust MPC: Enforces constraints for all possible realizations within a bounded uncertainty set. Guarantees hard safety but can be overly pessimistic, leading to poor performance.
• Stochastic MPC with Chance Constraints: Allows for rare, tolerable violations (e.g., 1% probability). This trade-off between risk and performance is quantifiable and often more practical for systems where absolute worst-case events are extremely rare.
• Design Choice: Selecting the violation probability ε is a fundamental engineering trade-off between performance aggressiveness and safety margin.

Stochastic Model Predictive Control (SMPC) is an advanced control framework that explicitly accounts for probabilistic uncertainties—such as random disturbances or noisy measurements—within its optimization. Unlike deterministic MPC, which uses a single nominal prediction, SMPC propagates uncertainty through the system model. Its primary goal is to compute control actions that optimize expected performance while satisfying constraints in a probabilistic sense, often using chance constraints. This provides a less conservative and more realistic alternative to worst-case Robust MPC.

Robust MPC is a control strategy designed to guarantee constraint satisfaction and stability for all possible realizations of bounded model uncertainties and disturbances within a predefined set. It employs a worst-case analysis, leading to potentially conservative control actions. This contrasts with Stochastic MPC, which uses chance constraints to allow for a small, specified probability of constraint violation, often resulting in better average performance. Key techniques in Robust MPC include:

• Constraint tightening, where the feasible set is shrunk to create a safety margin.
• Tube-based MPC, where controllers are designed to keep the system state within a bounded "tube" around a nominal trajectory.

An Optimal Control Problem (OCP) is the foundational mathematical optimization solved at each control step in MPC. It is defined by:

• A dynamic model (e.g., differential/difference equations) predicting state evolution.
• A cost function to be minimized (e.g., tracking error, control effort).
• A set of constraints on states and control inputs. When chance constraints are incorporated, the standard OCP becomes a Stochastic Optimal Control Problem. The constraints are no longer deterministic inequalities but probabilistic statements, such as Pr(state ≤ limit) ≥ 0.95. Solving this typically requires reformulation into a tractable deterministic approximation.

Constraint handling refers to the methods by which an MPC controller enforces limits on system variables.

• Hard Constraints are absolute physical limits (e.g., actuator saturation, collision avoidance) that must never be violated. Chance constraints provide a probabilistic interpretation of hard constraints under uncertainty.
• Soft Constraints are desirable limits (e.g., staying within a preferred zone) that can be temporarily violated at a penalty. They are implemented using slack variables added to the cost function to ensure the optimization problem remains feasible. In practice, critical safety limits are often treated as hard or chance constraints, while operational limits are softened.

A Control Barrier Function (CBF) is a mathematical tool used to enforce safety by defining a forward-invariant safe set. If a system starts inside this set, a controller designed with a CBF will guarantee it remains inside for all future time. While traditionally used in real-time safety filters, CBFs can be integrated with MPC:

• As a predictive safety filter, where MPC provides a desired performance-oriented input, and a CBF-based module minimally modifies it to ensure instantaneous safety.
• As additional constraints within the MPC OCP to enforce safety over the prediction horizon. This combines the long-term optimization of MPC with the rigorous, constraint-based safety guarantees of CBFs.

Scenario-based optimization is a prominent method for handling chance constraints in Stochastic MPC. Instead of dealing with complex probability distributions analytically, it approximates the probabilistic constraint by drawing a finite number N of random scenarios (sample trajectories) of the uncertain parameters. The optimization is then performed such that constraints are satisfied for a high percentage of these sampled scenarios. This approach:

• Converts a stochastic problem into a larger, deterministic one.
• Provides probabilistic guarantees on constraint satisfaction that improve with the number of scenarios.
• Is computationally intensive but highly flexible, as it does not require specific assumptions about the uncertainty distribution (beyond the ability to sample from it).