In the design of autonomous systems, we typically ask two different types of questions. The first is qualitative: "Will the robot reach its destination without crashing?" This is a binary, logical concern—a matter of success or failure. The second is quantitative: "How much battery will the robot use?" This is a numerical concern—a matter of degree and optimization.
Historically, these two questions were treated as separate disciplines. Qualitative properties were the domain of formal methods and temporal logic, while quantitative properties belonged to control theory and optimization. However, as Rajeev Alur, Suguman Bansal, and Shaull Almagor argue in Equilibria in Quantitative Concurrent Games, real-world systems must satisfy both simultaneously. A robot that never crashes but runs out of power halfway to its goal is just as useless as a fast robot that hits a wall.
The emerging research direction in reactive synthesis aims to unify these disparate objectives into a single mathematical framework, allowing designers to automatically construct controllers that are guaranteed to be both safe and efficient.
The Mathematical Gap
The difficulty in combining these objectives stems from their underlying mathematical structures. Qualitative goals are often expressed in Linear Temporal Logic (LTL) and solved using automata-theoretic approaches. Quantitative goals, such as maximizing a "mean-payoff" (the average reward over time), are typically solved using Linear Programming or dynamic programming in Markov Decision Processes (MDPs).
When you try to solve them together, the state space explodes. As Mickaël Randour, Jean-François Raskin, and Krishnendu Chatterjee point out in Strategy Synthesis for Multi-Dimensional Quantitative Objectives, integrating these disparate types requires different mathematical treatments that don't always play well together, especially when dealing with multi-dimensional constraints like simultaneous limits on energy, time, and risk.
Automata as a Universal Language
The breakthrough in this field involves finding a "lingua franca" that can represent both logic and numbers. One powerful approach is to translate logical specifications into specific types of automata that can be embedded directly into the numerical optimization process.
In LTL-Constrained Steady-State Policy Synthesis, Jan Kretinsky demonstrates how to reduce heterogeneous requirements—qualitative LTL, quantitative steady-state frequencies, and average rewards—into a single Linear Program. This is achieved using Limit-Deterministic Büchi Automata (LDBA). By converting the logical "safety" requirement into an LDBA, the synthesis tool can treat the logical state as just another variable in the MDP, allowing the system to optimize for rewards while strictly adhering to the logical constraints.
From Optimization to Satisficing
Another key intuition is moving away from pure optimization toward satisficing. In On Satisficing in Quantitative Games, Krishnendu Chatterjee, Moshe Y. Vardi, and Suguman Bansal suggest that instead of finding the absolute best numerical value, we should treat quantitative thresholds (e.g., "use less than 10kJ of energy") as logical constraints.
By representing these numerical thresholds as automata, they can be composed with LTL specifications via simple automata intersection. This modularity allows a system to treat "staying under a budget" with the same formal rigor as "avoiding an obstacle." This approach simplifies the synthesis of strategies that must respect resource constraints while satisfying complex reachability properties.
Handling Adversarial Environments
In many scenarios, systems don't operate in a vacuum; they operate in environments that may be uncertain or even adversarial. Wei Zhao and colleagues in A synthesis method for zero-sum mean-payoff asynchronous probabilistic games extend this unification to game-theoretic settings. They weight the expected reward (quantitative) by the probability of satisfying the underlying logical specification (qualitative). This ensures that the synthesized system doesn't just chase high rewards in paths where it is likely to violate safety rules.
Similarly, Nir Piterman explores "combined winning" in Combinations of Qualitative Winning for Stochastic Parity Games, providing a framework for systems that must balance absolute "sure winning" for critical safety with "almost-sure winning" for general performance goals.
Practical Implementation: MultiGain
The theoretical unification of these objectives has led to practical tools like MultiGain 2.0, developed by Severin Bals, Jan Kretinsky, and others. As described in MULTIGAIN 2.0: MDP controller synthesis for multiple mean-payoff, LTL and steady-state constraints, this tool allows designers to visualize the trade-offs between competing objectives using Pareto curves.
This highlights the ultimate goal of this research: providing a unified "behavioral framework," as envisioned by Arjan van der Schaft in A behavioral framework for compositionality, where safety, performance, and efficiency are not treated as isolated problems, but as facets of a single, coherent system behavior.