Heuristics for Solver-Aware Systems Architecting: A Reinforcement Learning Approach

Abstract The crowdsourcing literature has shown that domain experts are not always the best solvers for complex system design problems. Under certain conditions, novices and specialists in adjacent domains can provide novel solutions at lower costs. Additionally, the best types of solvers for different problems are dependent on the architecture of complex systems. The joint consideration of solver assignment and system decomposition, referred to as solver-aware system architecting (SASA), expands traditional system architecting practices by considering solver characteristics and contractual incentive mechanisms in the design process and aims to improve complex system design and innovation by leveraging the strengths of domain experts, crowds, and specialists for different parts of the problem. The joint consideration of problem decomposition and solver assignment decisions in SASA renders the design space exponentially more complex. Therefore, new computationally efficient and mathematically rigorous methods are needed to explore this high-dimensional space and extract reliable heuristics. To address this need, this paper presents a computational approach using a Markov decision process (MDP) formulation, Q-learning, and Gaussian mixture models. Together, these techniques explore the large space of possible solver–module assignments by modeling the sequential nature of solver assignment decisions, capturing these temporal dependencies, thereby enabling optimization for long-term expected rewards, and analyzing reward distributions. The approach identifies heuristics for solver assignment based on the designer’s preference for cost-performance trade-off through the parameterized reward function. The approach is demonstrated using a simple and idealized golf problem, which has characteristics similar to design problems, including how the problem is decomposed into interdependent modules and can be solved by different solvers with different strengths that interact with the module type. The results show that the proposed approach effectively elicits a rich set of heuristics applicable in various contexts for the golf problem and can be extended to more complex systems design problems.