NCCS lab members are involved in a variety of research projects, including the following.

Complexity in Human Teams

Human cognition can be viewed as a non-linear dynamical system receiving inputs and produces outputs that vary over time. Chaos theory describes the behavior of certain non-linear dynamic systems that may appear to proceed according to chance even though their behavior is determined by precise laws and is highly sensitive to initial conditions.  Tom Clarke and his research group are studying the effectiveness of using chaos theory to measure and evaluate human and robot performance, but the approximations used are performed off-line due to their data and computational time requirements.

The NCCS is assisting in this research by exploring ways in which transfer learning can be used to make the computational effort of these measures more manageable in real time.  Our approach is to use traditional machine learning methods on existing data in order to determine general baseline parameters that yield fast and accurate approximations, then tune these parameters in real time on specific, novel sequences in order to improve computational speed.  The goal is to be able to produce timely, on-line performance feedback for agents in the system.

Learning Robust Behaviors in Mixed-Agent, Heterogeneous Teams

Once humans are a part of a multiagent team, many issues become much more complex.  Machine learning methods for multiagent systems typically employ simulations for some portion of the learning process; however, in mixed-agent teams (teams with both humans and machine agents) simulated agents will not simply include those over which learning methods have control (”intrinsic” team members) but also include agents that learn and behave in ways external to, and are more complex than, their software-based counterparts (”extrinsic” team members).  Here optimality is far less important in many cases than robustness—-autonomous team members must behave in ways that lead to sufficient team performance even when some team members behave in unexpected ways.  Our study focuses on two key questions:  How can one effectively model and simulate extrinsic team members to aid in the learning process?  And what are effective mechanisms to evaluate mixed-agent team performance and learn intrinsic agent behaviors?

We consider several methods for modeling extrinsic member behaviors at different levels (e.g., neural networks, Bayesian belief nets).  The object is not to produce faithful models of how a human team members behave, rather to produce task-oriented behaviors that vary in abstraction and quality.  The models may be adjusted by hand or by software, but this process will be independent of the machine learning process discussed below.

The behaviors of the intrinsic team members employ natural computation based methods involving distributed and decentralized control mechanisms, physicomimmetics.  High-level interaction models are designed modularly by hand, but the specific parameters encoding the behaviors are developed by coadaptive learning methods. Early theoretical research into certain types of compositional coevolutionary algorithms indicates that they are suited for producing robust solutions.  Our project will deepen this foundation by clarifying mixed-agent performance goals that focus on robustness, developing theoretical and empirical measures for evaluating these robustness performance goals, and outlining those aspects of coadaptive learning methods that are likely to optimize such measures.

Evolution and Evolvability

More details soon …  

Design Methods for Heterogeneous, Modular Swarms

Along with the Adaptive Systems section at the NCARAI and the Distributed Robotics lab at the University of Wyoming, we have been involved in developing a principled method for constructing modular, heterogeneous swarms.  The NCCS lab uses this method for developing high-level modular designs of multiagent teams.

The method generalizes Dr. Spears’ artificial physics based technique for representing swarm behaviors by providing explicit notions of agent and interaction specialization. Additionally, our method uses a novel graph-based design tool for swarm-based behaviors for multiagent teams. This method includes engineer-provided knowledge through explicit design decisions pertaining to specialization, heterogeneity, and modularity. The representational power of our generalized representation and design methods has been demonstrated on a variety of problems, including a well-known multiagent resource protection problem.  Using our approach, it is straight forward to construct modular designs by hand, resulting in scalable and intuitive heterogeneous solutions.

Decomposition in Coadaptive Learning Systems

Coadaptive learning systems involve decomposition by definition:  concurrent learning algorithms working on different aspects of a problem, game, or simulation imply (explicitly or implicitly) a division in representation of some kind.  For many such systems, this partitioning is performed explicitly and a priori (e.g., traditional compositional coevolutionary approaches). The representational issues surrounding how to choose to decompose a problem for the successful application of coadaptive learning approaches are complex and often quite counter intuitive.  For example, the property of linear separability is neither a sufficient nor necessary condition for success.    

Though some of these issues have been addressed for specific classes of algorithms, such as the afore mentioned compositional approach, we believe many of them are general to a variety of coadpative methods.  Our approach is to explore decompositional issues from both theoretical and empirical points of view — to answer such questions as: What properties of the problem affect representational choices with respect to decomposition in coadaptive learning algorithms?   How do other algorithmic design choices  relate to the question of decomposition?  What are the tradeoffs between optimal decompositions and intuitive decompositions?