Towards a Distributed Multiagent Learning-Based Design Optimization Method

Complex engineered systems create many design challenges for engineers and organizations because of the interactions between subsystems and desire for optimality. In some conceptual-level optimizations, the design problem is simplified to consider the most important variables in an all-in-one optimization framework. This work introduces a stochastic optimization method which uses a distributed multiagent design method in which action-value based learning agents make individual design choices for each component. These agents use a probabilistic action-selection strategy based on the learned objective values of each action. This distributed multiagent system is applied to a simple quadrotor optimization problem in an all-in-one optimization framework, and compared with the performance of centralized methods. Results show the multiagent system is capable of finding comparable designs to centralized methods in a similar amount of computational time. This demonstrates the potential merit of a multiagent approach for complex systems design.