In real-world engineering, we rarely have the luxury of optimizing for a single metric. A drone cannot simply maximize its speed; it must also conserve battery power. A neural network cannot blindly chase accuracy; it must operate within the latency limits of its target hardware.
When multiple objectives conflict, there is no single "perfect" solution. Instead, there is a Pareto frontier—a boundary of optimal trade-offs where improving one metric forces the degradation of another.
To navigate this frontier, researchers compress multi-dimensional goals into a single, tunable mathematical landscape. This approach, known as scalarized multi-objective optimization, allows systems to dynamically balance competing priorities by adjusting scalar weights.
Collapsing Dimensions via Scalarization
The core challenge of multi-objective optimization is that standard optimization algorithms—such as gradient descent, linear programming, or reinforcement learning—require a scalar value to determine which direction is "up." They cannot natively handle a vector of competing goals.
Scalarization solves this by combining disparate objectives into a single cost function using weighted coefficients. In its simplest form, a linear weighted sum combines two objectives, $f_1(x)$ and $f_2(x)$, into a unified objective $F(x)$:
$$F(x) = \alpha \cdot f_1(x) + (1 - \alpha) \cdot f_2(x)$$
By changing the scalar weight $\alpha$ between 0 and 1, the optimization landscape warps, shifting the global minimum along the Pareto frontier. This mechanism transforms a complex multi-dimensional trade-off into a familiar single-objective problem that existing solvers can traverse.
For example, in cloud computing infrastructure, Thai et al. apply this concept to task scheduling in Executing Bag of Distributed Tasks on the Cloud: Investigating the Trade-offs Between Performance and Cost. They introduce a tunable preference parameter to balance execution speed against monetary expenditure. This scalar value guides a heuristic algorithm, scoring candidate deployment plans based on the user's specific budget-performance preference.
Objective B (e.g., Latency)
^
| * (High Alpha: Prioritizes Objective A)
| \
| \ Pareto Frontier
| \
| * (Low Alpha: Prioritizes Objective B)
+-----------------------------------> Objective A (e.g., Energy Cost)
This scalarization trick is equally powerful in deep learning. In Dynamic Deep Neural Networks: Optimizing Accuracy-Efficiency Trade-Offs by Selective Execution, Deng et al. design a neural network that decides on-the-fly which layers to execute based on the complexity of the input. By embedding a tunable hyperparameter directly into the reinforcement learning reward function, they weight classification accuracy against computational cost. A single training framework can thus generate a variety of routing behaviors tailored to different hardware constraints.
Dynamic Adaptation to Real-Time Constraints
The true utility of a weighted objective function is not just finding a single compromise, but the ability to shift along the Pareto frontier dynamically as environmental conditions change. Instead of hardcoding a static policy, operators can turn a single "knob" to reconfigure a system in real time.
In industrial IoT networks, communication bandwidth and device battery levels fluctuate constantly. Khan et al. demonstrate this adaptability in Resource Optimized Federated Learning-Enabled Cognitive Internet of Things for Smart Industries. By utilizing adjustable weighting parameters within their optimization framework, they allow edge devices to dynamically prioritize either energy conservation or processing speed based on real-time hardware states. When a device's battery is low, the weight shifts to favor energy preservation; when time-critical diagnostics are required, the weight shifts to maximize throughput.
Physical systems operating at high speeds rely on this exact tunability for safety. In Aggressive Online Control of a Quadrotor via Deep Network Representations of Optimality Principles, Li et al. use a parameterized cost function to train an onboard neural network for drone flight. The weighting factor balances aggressive, time-optimal maneuvers against conservative, power-saving trajectories. Because the policy is parameterized by this weight, the drone can alter its flight profile mid-air to adapt to changing mission parameters or battery levels.
This framework also extends to environmental sustainability. In Carbon-Aware EV Charging, Cheng et al. introduce a tunable weighting parameter to balance carbon emissions against total energy delivery for electric vehicle charging. This allows charging stations to automatically defer power draw to periods of high renewable energy generation, trading off minor delays in charging completion for significant reductions in grid carbon footprint.
Beyond Linear Combinations: Non-Convex Frontiers
While linear weighted sums are mathematically straightforward, they suffer from a fundamental limitation: they struggle with non-convex Pareto frontiers. If the optimal trade-off curve bends inward, a linear weight vector will "jump" over the concave region, rendering those compromise solutions unreachable no matter how the weights are adjusted.
To overcome this, researchers employ advanced scalarization techniques like compromise programming and Tchebycheff optimization, which use min-max formulations to guarantee coverage of the entire frontier.
Linear Scalarization (Fails on Concave Fronts) Tchebycheff Optimization (Succeeds)
^ ^
| \ | | /
| \ (Jumps over the gap) | | / (Contours bend around
| \ | |/ the non-convex corner)
+-------------------> +------------------->
For instance, designing edge server networks involves highly non-linear relationships between link construction costs and network performance. In Cost-Effective Edge Server Network Design in Mobile Edge Computing Environment, Jin et al. leverage compromise programming. This mathematical formulation measures the distance to an ideal, unfeasible reference point, allowing network designers to stably balance construction budgets and latency targets even within non-convex design spaces.
A similar challenge arises in 6G wireless networks featuring Integrated Sensing and Communication (ISAC), where the same waveform must simultaneously act as a radar sensor and a data transmitter. As shown by Wei in Waveform Design for MIMO-OFDM Integrated Sensing and Communication System: An Information Theoretical Approach, resources must be carefully split between these two fundamentally different tasks using information-theoretic metrics.
To handle the complex trade-offs in robotic IoT environments, Ni et al. employ a weighted Tchebycheff optimization approach in Reconfigurable Intelligent Surface for Internet of Robotic Things. By minimizing the weighted maximum deviation from individual targets rather than summing them, the system successfully maps out the non-convex trade-off space between sensing precision and communication throughput, ensuring both objectives are maximized in tandem.