Pareto-based multi-objective methods can be expensive, and it isn't clear how they extend to multiagent systems. I developed an adaptive scalarization method that transforms the objective space to be easier to deal with. With this a cheap linear combination of transformed objectives performs on-par with expensive Pareto-based methods. Results include a 10x reduction in computation time and the ability to attain the same performance as other state-of-the-art methods in less than 1/3 of the samples. We have applied PaCcET to domains such as rover exploration, hybrid power generation, and microgrid coordination.
For more details on the PaCcET algorithm, see my 2014a SEAL paper.
Autonomous rovers are one of the most exciting opportunities for exploring other planets. A team of ten robots can perform much better on this task than a single robot running ten times as long, so multiagent systems are a natural solution.
One domain I use to test algorithms I develop models multiple robots exploring an unknown area, attempting to find points of interest. They can have heterogeneous sensor packages or communication capabilities, and must cooperate to make the best observations of their environment as quickly as possible.
My research focuses on the optimization of messy, difficult to model systems. In this type of system, it is typically not a simple task to enumerate the quantity that should be optimized, which leads to a multi-objective optimization (MOO) problem. Furthermore, as we speak of increasingly complex systems, the benefits of decentralization become greater, which creates a multiagent system (MAS). My research is at the interface of these two types of problems (which are each fields of research in their own right), creating a multi-objective multiagent systems (MOMAS).
More details on my research can be found in this one-page white paper.
Air traffic is an extremely complex system, and unforeseen consequences can arise from the interaction between humans and new technologies. I developed an agent-based simulation to assess the possible impact of introducing new technologies into air traffic control.
For more details, see my 2014 GECCO paper.
Assigning credit to individual agents for the team's successes or failures is a key problem for a multiagent systems. How do we do this for a multiple objective problem? This question in a broad sense forms the central point of my doctoral dissertation. I partially answered this question in this 2014b SEAL paper, where I studied credit assignment using objective aggregation methods.