AI links: Organizations · Resources · Journals · Conferences · People · RL
Other links: Books · Libraries · Music · News & Info · Search · Software · Travel ·Amsterdam · Boston

Edwin D. de Jong

LDD Group, ICS, Universiteit Utrecht

Publications·Research·Thesis·Contact details·CV



NOTE for prospective students, peer review requests, etc.:
As of 2008 I've been working full-time for an innovative data-mining company called Adapticon (www.adapticon.com).
While I'm still affiliated with Utrecht University, I'm currently unable to accept any requests for internships, program committee memberships, etc. Thanks for your understanding.

Activities:

Publications

To view or print these online publications, use a postscript or PDF viewer. Please feel free to contact me if you experience any difficulties in viewing a paper, or of course if you have questions or comments about the contents of one of the papers. See here for a list including abstracts.

Book chapters

Refereed Journal

Bioinformatics Abstracts

Refereed Conferences and Workshops

Technical Reports

Research Interests

Coevolution
In most methods of evolutionary computation, a fixed fitness function is defined. This results in a particular gradient, and determines the course of evolution beforehand. If on the other hand the evaluation of individuals is based on interactions with other evolving individuals, then in principle a potential for open-ended evolution can arise. Evolutionary methods with this form of evaluation are called co-evolution. Coevolution is substantially different from conventional evolution, in that the ranking of individuals is dependent on the co-evolving individuals. An interesting recent development in coevolution is Pareto-coevolution, which views the interactions with other evolving entities as objectives, in the sense employed by Evolutionary Multi-Objective Optimization (EMOO).

Modularity
If the design of complex systems is viewed as a process of arranging large numbers of primitive elements in a specific manner, then the space of possible designs is prohibitively large. However, if some choices concerning aspects of a design can be made independent of other choices, then the required size of the effective search space can be greatly reduced by recursively combining such independently made choices. A substantial reduction is still possible if choices are not independent, but if the dependencies are merely reduced compared to the case where each choice depends on all other choices. Problems exhibiting this property are called modular, and the principle of modularity is seen in a wide variety of design processes, ranging from nature to engineering and from music composition to software design. I am interested in seeing how it can be achieved by evolutionary processes. A particular direction I am exploring is to view the evolution of modules and the assembly thereof as a process of coevolution.

PhD Thesis

My PhD thesis titled Autonomous Formation of Concepts and Communication is available here.

Research History

After receiving the MSc from Delft University of Technology, I went to the VUB AI Lab in Brussels, Belgium, headed by Prof. Luc Steels. The first topic I've worked on is coordination between software agents [3]. The idea behind this research is that if agents are to coordinate their actions with other agents whose competences are unknown in advance, they can achieve this coordination by sending evaluations to each other. This turned out to be a feasible idea [6], and led me to the interesting field of reinforcement learning. My research in that area included the exploration/exploitation dilemma [4], and generalization [5].

These earlier efforts led me to research concerning the development of language. Concepts are not assumed to be present in the agents already, but are developed based on interaction with the environment. A first result was that a shared lexicon can emerge in a community of agents that independently form private concepts about their environment [7]. As a next step, I have suggested the principle of autonomous concept formation, using which agents can form situation concepts [9]. A situation concept is a set of states of the environment that is relevant for an agent to distinguish from other possible states because the situation is predictive of future states. When agents build up a lexicon relating situations to words, they can use this lexicon to communicate information about the environment. Thus, uncertainty due to incomplete perception can be overcome by means of communication. In subsequent work, I investigated communication as a dynamical system in which associations between concepts and words continuously change [10], and perfect systems of communication are point attractors [11]. Details can be found in my PhD thesis.

Next, in August 2000, I went to Boston to do a two-year postdoc at Jordan Pollack's DEMO Lab. After these stays abroad, I'm back again in my home country: The Netherlands.

Organizations: AAAI, IEEE, ISAB, BNVKI, Evonet, Emergentia.

Conference Reports

  • GECCO'03 conference report. The Genetic and Evolutionary Computation Conference. BNVKI Newsletter vol.~20, no.~4, 84-87.

  • IJCAI'99 conference report. The Sixteenth International Joint Conference on Artificial Intelligence. BNVKI Newsletter vol.~16, no.~5, 132-135.

  • EvoIASP'99 conference report. The First European Workshop on Evolutionary Image Analysis and Signal Processing. BNVKI Newsletter vol. 16, no. 3, pp. 74-75. A part of this article also appeared in: EvoNews. Newsletter of EvoNet - The Network of Excellence in Evolutionary Computing. Issue 11, Summer 199, p.12

  • SAB'98 conference report. The International Conference on Simulation of Adaptive Behavior. NVKI Newsletter vol. 15, no. 4., pp. 112-116.

    Contact information:

    See note above. E-Mail: d e j o n g @ c s . u u . n l