Section: New Results
Developmental Representations for Evolutionary Design
Developmental representations provide a powerful framework for the automatic design of complex structures. In this setup, referred to as artificial ontogeny, global effects are reached through local interactions of elements (or cells), with some guarantees about the scalability and robustness of the solution structures. Proofs of principle of the approach, concerning the design of bridges or multi-robot cooperation, have been given in Alexandre Devert's PhD (defended in May 2009), including the evolutionary optimization of Echo State Networks using new fitness objectives and the design of a specific stopping criterion. This stopping criterion not only yields significant computational savings; it also provides design solutions which are outstandingly robust w.r.t. environmental perturbations. A prototypical application, the design of a truss structure, features a single cell template used all over the structure, flexibly achieving different behaviours depending on the local environment, and resulting in a low-dimensional search space.
Self-Adaptive and Swarm Robotics
In the European Symbrion IP (2008-2012 – http://www.symbrion.eu/ ), TAO's role is to provide the robotic swarm with learning and evolution facilities, meeting the constraints of in-situ and on-board design. During his 3-month visit at the Vrije Universiteit, Amsterdam, Nicolas Bredèche investigated the on-board evolution of autonomous robots  ,  . Position papers regarding the self-organization of robotic swarms have been published  ,  and Jean-Marc Montanier started his PhD in Sept. 2009 on this topic. Interestingly, the Complex Systems and the Large and Deep Networks SIGs strongly interact to investigate the computational properties of e.g., Echo State Networks from different and complementary perspectives. The good properties of ESNs to achieve autonomous robotic control for non-markovian tasks, e.g. related to the Tolman maze, have been shown in Cedric Hartland's PhD, defended in Nov. 2009  ,  .
Data Mining and Information Theory for Autonomous Robotics
In collaboration with the University of Kyushu, the joint SyDiNMaLaS ANR-JST proposal concerns the principled use of the robotic logs to devise and debug a robotic controller (2008-2011). A first step along this line investigates the on-board evolutionary optimization of the controller entropy, computed from its trajectory  and Pierre Delarboulas started his PhD in Sept. 2009 on this topic.
TAO's involvement in Robotics has been made visible through the co-organization of workshops collocated with the mainstream robotic IROS conference  (in collaboration with ISIR, Paris 6) and the European conference on Machine Learning and Data Mining (in collaboration with U. of Kyushu).
A mailing list on Evolutionary Robotics (”evoderob”, managed by N. Bredeche, S. Doncieux and J.B. Mouret) involve most researchers from the ER community. This activity has been widely disseminated  , Fête de la science .
Social Systems Modelling
Multi-agent systems have been investigated to simulate, understand and optimize complex systems, and more specifically social systems. Multi-agent based simulation (MABS) provides a fast prototyping framework for modelling agent behavior, interaction rules and scenarii, while enabling the multi-scale observation of the system.
Our 2008 study regarding the French Academic Labor Market was resumed to account for the University learning behaviour  , showing that e.g. good but not top-ranked universities learn to lower their expectations to compensate for the market saturation. Another complex, thrust-based market is the Rungis wholesale market; a thorough model thereof was proposed in collaboration with University of Coimbra and AUDENCIA school of management  ; the thrust impact on the maximization of both the profit and the exchange volume has been highlighted and confronted to empirical observations  .
Independently, new models of cognitive agents relying on polychronous networks and Memory Evolutive System have been proposed in collaboration with University of Sao Paolo  ).
Lastly, Multi-agent systems were used to study decentralized coalition formation and restructuration protocol in a multi-objective framework. A proof of principle of the approach, delivering Pareto optimal solutions in a small-size class-scheduling problem has been proposed in collaboration with Université Lyon-1 and Université Paris Dauphine  .
A Statistical Physics Perspective
Basic tools from statistical physics (scaling, mean-field techniques and associated distributed algorithms, exactly-solvable models) and probability have been used to model and optimize complex systems, either standalone or combined with MABS approaches.
Data streaming for Autonomic Computing (section 6.1 ) has been identified as a promising field of applications; the message passing Affinity Propagation algorithm has been extended using on-line aggregation and hierarchical processing of the data stream  ; the scaling analysis based on a renormalization-based approach yields an almost costless way of finding the “true” number of clusters in the data  .
Another message-passing approach has been investigated to model road traffic in the context of the ANR Travesti project, aimed at learning an approximate Markov Random Field based on generalized Bethe free energy approximations, in collaboration with the Imara INRIA project. In  we have obtained a surrogate solution which can efficiently encode a supperposition of patterns in the form of Belief-Propagation fixed points; this solution provably asymptotically behaves like an unsupervised Hopfield model. Work in progress concerns the design of exactly solvable models relevant to the understanding of the fundamental diagram of traffic flow in the ANR Travesti context; independently, the design of a message passing algorithm for sampling the Pareto Front of a multi-objective combinatorial optimization problem has been considered in the STREP Gennetec context.
Sequential Representation for Temporal Planning
On-going collaboration with Thalès (Jacques Bibaï's CIFRE PhD; DESCARWIN ANR project, accepted in 2009) is concerned with formalizing, validating and improving the Divide-and-Evolve (DAE) approach to Temporal Planning. Within IPC'08 (International Planning Competition), DAE was found to fail on 50% of the problem instances, while matching or improving the best known results on the other instances  . Based on the analysis of those results, several improvements of DAE have been designed. Among the most promising ones are the use of heuristic bounds on the earliest possible time where an atom can become true  , the tuning of the (many) algorithm parameters through a racing procedure  , and the combination with different (non-optimal) embedded planners, that seems to paradoxically improve the results a lot  .