UC Berkeley

Recently, the artificial intelligence has achieved a significant success with applications in various domains including transportation, smart building, robotics, economy and so on. More and more traditional system entities have been entitled with full or partial autonomy, allowing them to make their own decisions and moves based on the specific surrounding environments. An integration of multiple such intelligent entities is called a multi-agent system (MAS) where the agents need to interact with each other effectively and efficiently to attain cooperation and optimal system performance. As to fulfill this more challenging intelligent interaction objective, the traditional control approaches will not suffice and more advanced algorithms become essential.

In this dissertation, three system structures for interactive control systems, centralized, distributed and decentralized, are discussed with application in intelligent building and autonomous driving. Several concrete interactive control algorithms are proposed and verified.

In the centralized control system, a single central agent with the whole system information available is in charge of making decisions for all the agents. The systemwise cooperation solution is thus directly obtained and all the interactions involved are optimally addressed. Chapter 3 and 4 adopt such centralized control strategy for the intelligent building system. In order to save energy consumption and satisfy the occupants' thermal comfort demand, a combination of feedforward iterative learning control (ILC) and iteratively tuned feedback controller is designed to compensate both repetitive and non-repetitive disturbance components. Chapter 3 proposes an iterative controller design algorithm via optimization solving and stabilizing feedback projection. In Chapter 4, the concurrent design of feedforward ILC and causal stabilizing feedback controller is introduced, where both controllers are simultaneously solved by one optimization.

However, the centralized approach's complexity grows with the problem size, which leads to failure for large-scale systems. The distributed control strategy is introduced as an alternative for such high-dimensional control problems. In the distributed system, a communication network enables the information exchange among agents. Therefore, each agent can keep broadcasting and updating its local controller until a convergence to the cooperative solution is reached. In Chapter 5, a distributed cooperative controller design method is developed for intelligent building thermal control with convergence property theoretically proven.

For a system with no global communication, agents of which follow different control policies, the decentralized control structure is the only valid solution, where each agent designs its local controller independently based on estimated information of others. In Part II of the dissertation, several decentralized interactive control algorithms are proposed for the autonomous driving system. In Chapter 6, an optimization-based negotiation with both concession and persuasion is formulated for vehicle agent's decision making in various interactive scenarios. A Bayesian persuasion based algorithm for interactive driving is explored in Chapter 7. In the algorithm, the ego vehicle agent (persuader) intends to manipulate the interacting vehicle agent (information receiver)'s belief about the current driving situation via observable driving behavior. In Chapter 8, the interaction between two vehicle agents is defined as a two-player persuasion game, the mixed Nash equilibrium of which denotes the agents' optimal intention probabilities. The optimal intention is then expressed via the ego vehicle's driving trajectory planned by an optimization with the intention expression constraint.