Optimizing for multiple objectives at once is a complicated challenge… especially when the environment changes over time.
Authored by LYU Guodong
From navigating our personal priorities to determining business strategy at the highest level, many tasks require us to balance competing goals. In some cases, this optimization is fairly straightforward: When one goal is clearly more important than the other, and when the environment stays static over time, it’s easy enough to see what to do. But in other cases, the importance of different goals may vary, or be difficult to determine. What does it take to balance competing goals in such a dynamic, opaque environment?
Our recent research explored this critical question in the context of an industry that’s tasked with balancing numerous conflicting goals every day: ride-sharing platforms. Ride-share companies like Uber, Lyft, and Didi have developed complex algorithms to decide which passengers to match with which drivers in real time. In an ideal world, these matching algorithms would balance short-term considerations such as maximizing total revenue and minimizing passenger pick-up distances with long-term goals such as fostering a reliable community of drivers and ensuring consistently high service quality for passengers. But in practice, we’ve found that many commonly-used algorithms simply assign passengers to the nearest available driver, to reduce wait times and increase the chances that the match is accepted by both parties.
Reliably Balancing Conflicting Goals Isn’t Easy
Part of the reason for this typical strategy is that balancing the conflicting goals of revenue gain, service excellence, and operational efficiency in the dynamic context of ride-share matching is far from trivial. It’s not just a matter of short-termism — creating systems that can reliably optimize for multiple, non-static metrics is a complicated technical challenge.
That’s where our study comes in. Through our extensive research, we developed an innovative framework for real-time decision-making designed to help anyone strike a balance between conflicting objectives by dynamically adjusting how those different metrics are weighted. Here’s how it works:
A Dynamic Framework for Multi-Objective Optimization
Step 1: Establish Aspirational Targets
First, define your ideal target for each objective. For example, in the case of ride-share matching, what is your ideal revenue target, service quality score, and pick-up wait time? The ideal target can be designated as the utopia target, constructed by optimizing each objective function separately, yet unachievable simultaneously.
Step 2: Track Performance Gaps
Next, use automated tools to continuously measure the extent to which current outcomes deviate from these targets. In particular, we measure the performance gap between the achieved solution and the ideal target using a specific convex distance metric, such as the Euclidean distance function.
Step 3: Update Weight Functions
Finally, adaptively adjust the relative importance of each objective based on real-time tracking of the performance gaps. For example, if service quality starts to lag, the algorithm would temporarily prioritize dispatching higher-rated drivers. Conversely, if revenue begins to fall short, it would prioritize higher-value trips. We use the performance gaps, which are updated adaptively over time, as the priority weights to solve the multi-objective optimization problems. This approach guarantees that the achieved solution stays closest to the utopia target in the dynamic environment, achieving a delicate balance among the conflicting goals.
Inspired By Nature: The Gaze Heuristic
This framework was inspired by a natural, human phenomenon known as the Gaze Heuristic. The Gaze Heuristic is what allows people to automatically run in the right direction and at the right speed to catch a thrown object: Instinctively, when we see a moving object, we adjust our position such that the angle at which we have to look to keep the object in our sight remains the same. In other words, by continuously tracking the target and making real-time adjustments, we can optimize our decision-making in a complex and dynamic environment. Our framework applies the same principle to adjust any form of complex decision-making in response to shifting real-time data.
Of course, as with any theoretical framework, what really matters is how this approach performs in real life. As such, to validate the effectiveness our approach in practice, we ran a series of simulations and experiments using a ride-matching simulator developed by our industry collaborator, based on real-world data in several Chinese cities.
Validated By Real-World Data
Through these analyses, we found that our framework consistently and significantly outperformed conventional strategies that focus solely on minimizing pick-up distance. Specifically, real-world data from Beijing City indicated that our ride-matching system was associated with an improvement in average service quality of 0.92% per trip — which reflects more jobs being assigned to higher-quality drivers, which ultimately improves passengers’ experience — as well as a 1.88% increase in revenue per trip, which reflects that the algorithm prioritized longer trips that would bring in more revenue.
Importantly, these improvements did come at a cost: In our analysis, average pick-up distance per trip increased by 55.20 meters, and the rate at which ride requests were not fulfilled increased by 1.58%, which reduced revenue. However, taken together, these costs were outweighed by the benefits, as our approach still resulted in a 0.26% increase in net revenue. This suggests that overall, our framework optimized for near-term revenue more effectively than the legacy policy implemented in our industrial simulator, while also boosting both driver satisfaction and service quality (two factors that are critical for fostering sustainable growth and customer loyalty in the long term).
Clearly, our framework works well for ride-share matching. But what about other decision-making contexts? To further validate our findings, our framework was tested in two additional industries: food delivery and e-commerce.
Going Beyond Ride-Share
First, Meituan, a leading on-demand food delivery service in China, evaluated the effectiveness of our framework in balancing objectives such as courier efficiency and consumer experience throughout the food dispatch process. Through A/B tests comparing Meituan’s standard approach with our system for testing periods of at least two weeks each across five cities in China, Meituan’s team found that this framework led to an average improvement of on-time delivery rates of 0.33 percentage points, as well as a decrease in average travel distance of 0.92%.
And finally, we applied our framework to help craft online coupon assignment policies for an Indonesian e-commerce platform. In this context, we aimed to match coupon offers to customers in a way that would increase revenue while also maximizing total payment amounts, covering campaign costs, and adhering to a predetermined total budget. And again, simulations with real-world data suggested that using our approach to dynamically adjust how much these different factors were weighted in response to real- time data would help the platform boost revenue by more than 3% without exceeding its budget.
Balancing competing objectives is hard. Even with the best possible strategy, these trade-offs require compromise, and it’s not always obvious which factor to prioritize when. But with our simple, three-step framework, leaders in any industry can leverage real-time data to dynamically adjust their prioritization schemes in the moment, empowering organizations and decision-makers to optimize for revenue while ensuring that goals like sustainable success and user-centric innovation aren’t overlooked.
This article draws on the research paper titled "Multiobjective Stochastic Optimization: A Case of Real-Time Matching in Ride-Sourcing Markets," authored by Lyu Guodong, CHEUNG Wang Chi, TEO Chung-Piaw, and WANG Hai.
Lyu Guodong is an assistant professor of operations management at HKUST, with research interests in supply chain visibility, smart city operations, and sustainable development.