Gabriel Catalin Balan

MASON: A Multiagent Simulation Environment

MASON is a fast, easily extensible, discrete-event multi-agent simulation toolkit in Java, designed to serve as the basis for a wide range of multi-agent simulation tasks ranging from swarm robotics to machine learning to social complexity environments. MASON carefully delineates between model and visualization, allowing models to be dynamically detached from or attached to visualizers, and to change platforms mid-run. This paper describes the MASON system, its motivation, and its basic architectural design. It then compares MASON to related multi-agent libraries in the public domain, and discusses six applications of the system built over the past year which suggest its breadth of utility.

Tunably decentralized algorithms for cooperative target observation

Multi-agent problem domains may require distributed algorithms for a variety of reasons: local sensors, limitations of communication, and availability of distributed computational resources. In the absence of these constraints, centralized algorithms are often more efficient, simply because they are able to take advantage of more information. We introduce a variant of the cooperative target observation domain which is free of such constraints. We propose two algorithms, inspired by K-means clustering and hill-climbing respectively, which are scalable in degree of decentralization. Neither algorithm consistently outperforms the other across over all problem domain settings. Surprisingly, we find that hill-climbing is sensitive to degree of decentralization, while K-means is not. We also experiment with a combination of the two algorithms which draws strength from each.

Population implosion in genetic programming

With the exception of a small body of adaptive-parameter literature, evolutionary computation has traditionally favored keeping the population size constant through the course of the run. Unfortunately, genetic programming has an aging problem: for various reasons, late in the run the technique become less effective at optimization. Given a fixed number of evaluations, allocating many of them late in the run may thus not be a good strategy. In this paper we experiment with gradually decreasing the population size throughout a genetic programming run, in order to reallocate more evaluations to early generations. Our results show that over four problem domains and three different numbers of evaluations, decreasing the population size is always as good as, and frequently better than, various fixed-sized population strategies.

Long-term fairness with bounded worst-case losses

How does one repeatedly choose actions so as to be fairest to the multiple beneficiaries of those actions? We examine approaches to discovering sequences of actions for which the worst-off beneficiaries are treated maximally well, then secondarily the second-worst-off, and so on. We formulate the problem for the situation where the sequence of action choices continues forever; this problem may be reduced to a set of linear programs. We then extend the problem to situations where the game ends at some unknown finite time in the future. We demonstrate that an optimal solution is intractable, and present two good approximation algorithms.

Rapid development of large knowledge bases

This paper presents the Disciple-RKF methodology for rapid development of large knowledge bases which relies on importing ontological knowledge from existing knowledge repositories, on parallel development of separate knowledge bases by subject matter experts, and on the merging of these knowledge bases into a high performance integrated knowledge base. The paper discusses several issues related to ontology import and merging, and presents the results of a successful knowledge base development and integration experiment performed at the US Army War College.

Can good learners always compensate for poor learners

Can a good learner compensate for a poor learner when paired in a coordination game? Previous work presented an example where a special learning algorithm (FMQ) is capable of doing just that when paired with a specific less capable algorithm even in games which stump the poorer algorithm when paired with itself. We argue that this result is not general. We give a straightforward extension to the coordination game in which FMQ cannot compensate for the lesser algorithm. We also provide other problematic pairings, and argue that another high-quality algorithm cannot do so either.

A Demonstration of Neural Programming Applied to Non-Markovian Problems

Genetic programming may be seen as a recent incarnation of a long-held goal in evolutionary computation: to develop actual computational devices through evolutionary search. Genetic programming is particularly attractive because of the generality of its application, but it has rarely been used in environments requiring iteration, recursion, or internal state. In this paper we investigate a version of genetic programming developed originally by Astro Teller called neural programming. Neural programming has a cyclic graph representation which lends itself naturally to implicit internal state and recurrence, but previously has been used primarily for problems which do not need these features. In this paper we show a successful application of neural programming to various partially observable Markov decision processes, originally developed for the learning classifier system community, and which require the use of internal state and iteration.