Section: New Results

Understanding and mastering complex systems

Complex systems : simulation, control and definition

Adaptive control of a complex system based on its multi-agent model

Participants : Vincent Chevrier, Tomas Navarrete.

Laurent Ciarletta (Madynes team, LORIA) is an external collaborator.

We are interested in how to build a control mechanism for a complex/dynamic system. Specifically, we want to evaluate the effectiveness of creating a control mechanism based on a multi-agent model of the system[12] . Multi-agent models can be adapted to that purpose since usual approaches using analytical models as basis can be untractable when dealing with such systems; and because if we consider that the available control actions are meant to be applied locally, a multi-agent model is necessary. We are currently working on a case study within the dynamic networks domain, namely the free-riding phenomenon present in peer-to-peer networks.

We propose an architecture that gathers information from the system and uses it to parametrize and tune a set of multi-agent models. The outcome of simulations is used to decide which control actions have to be applied to the system, in order to achieve a predefined control objective. We consider that we do not have complete information to characterize the state of the system and hence would like to focus on the following two issues of the control problem that we have identified:

1. How to build a multi-agent model that represents the evolution of a dynamic network. That is, what to do when the information given by the simulation of the multi-agent is in contradiction with the information gathered from the system

2. How to build an adaptive control mechanism based on the multi-agent model of a dynamic network. That is, how to use the information given by the multi-agent model to achieve the control objective.

The architecture we proposed, is designed as a control loop composed of the following steps: estimate the state of the system and instantiate multi-agent models accordingly, simulate different control actions, choose a control action and apply it. From one cycle to another of the control loop, each step can be tuned (in terms of model parameters, control action selection process, sampling strategy, etc.) to overcome the previously mentioned issues of the control problem.

The architecture is currently specified in terms of a formal notation. We have already implemented the architecture within the context of the free-riding problem where we use the PeerSim simulator as the target system to control.

Within our case study, we have conducted two different sets of experiments to investigate under which conditions our control architecture can achieve its goal and to investigate the efficiency of different sampling methods to estimate the state of the network. We have effectively managed to drive the system to a state where the majority of the peers share, when the initial conditions, without intervention from our architecture, would drive the system to a state where no peer would share.

The elements of the architecture having an impact on the performance of the control obtained have been identified. These are: the initialization of the parameters of the models used to estimate the state of the system, predict the evolution of the system and test the possible control actions, as well as the strategy used to observe the system and the different time horizons to consider within the architecture.

The next steps are to better identify the advantages and limits of the proposed architecture and to widen the problem family in the free riding problem.

Multi Modeling and multi-simulation

Participants : Vincent Chevrier, Julien Siebert.

This work is undertaken in a joint Phd Thesis between MAIA and Madynes Team. Laurent Ciarletta (Madynes team, LORIA) is co-advisor of this PhD.

Complex systems generally require to use different points of view (abstraction levels) at the same time on the system in order to capture and to understand all the dynamics and the complexity. Being made of different interacting parts, a model of a complex system also requires simultaneously modeling and simulation (M&S) tools from different scientific fields.

Building a model and a simulation of a complex system from the interaction of the different existing M&S tools present in each scientific field involved, is also a complex task. To represent a complex system, we need to couple several models (multi-modeling) that each represents a part of the whole system. Each model could have been designed by and for a specific scientific domain. Making different models interact raises hard issues on model interoperability (semantic coherence, formalism compatibility). As many simulators exist in the scientific fields involved, a possible approach to make a simulation of a complex system is to reuse and to make interact these existing simulators. Since each simulator has been developed for specific purposes, making them interact (multi-simulation) raises simulation issues (interoperability, synchronization).

The multi-agent paradigm is an homogeneous solution both for multi-modeling and multi-simulation of complex systems. On the one hand, a multi-agent model per se is a multi-model: a multi-agent model is made of interacting agent models and environment models. On the other hand, agent oriented software engineering (AOSE) allows designers to create complex softwares as a set of autonomous, heterogeneous and interacting softwares (i.e. as a multiagent system). Robustness, scalability, openness, modularity and interoperability are some of the properties that AOSE allows to achieve.

This work explores the contribution of multiagent paradigm to the fields of multi-modeling and multi-simulation of complex systems.

The first contribution of this work is to propose an homogeneous multiagent meta-model (called AA4MM[4] ) that provides solutions both for multi-modeling and multi-simulation of complex systems by reusing existing and heterogeneous M&S tools . The core idea in AA4MM is to build a society of models, simulators and simulation softwares that solves the core challenges of multimodelling and simulation coupling in an homogeneous perspective. AA4MM has been implemented and used both for proof of concept and for a real case study. A proof of concept has been made by coupling different models together to develop a multi-model of a prey-predator model. This has permitted us to show both conceptual and operational properties of AA4MM such as interaction of heterogeneous models, modularity, interoperability.

This multiagent meta-model has been applied to model complex systems that are ubiquitous networks. Ubiquitous networks are highly dynamic computer networks that are composed of a great number of interacting and sometimes mobile nodes which can join or leave the system, interact together and where the environment plays a significant role either on radio communications or on the behavior of users. Modeling and simulation is the approach to evaluate these technologies or to build new ones.

Robustness of Cellular Automata and Reactive Multi-Agent Systems

Participants : Olivier Bouré, Vincent Chevrier, Nazim Fatès.

Our research on emergent collective behaviours focuses on robustness analysis, that is the behavioural resistance to perturbations in collective systems. We progressed in the knowledge of how to tackle this issue in the case of cellular automata (CA) and multi-agent systems (MAS).

We focused on the specific case of a perturbation of the updating scheme in CA, that is, changing the way cells are updated. Using similar ideas to the Influence-Reaction principle developed to resolve conflicts related to simultaneous actions, we created a new type of asynchronism, called beta-synchronism, which aims at disrupting the transmission of information about states between cells. We found out that the different types of asynchronism may induce radical change of behaviour for particular a value range of the synchrony rate [15] .

More recently, our interest focused on a bio-inspired discrete dynamical system. Using the formalism of a subclass of cellular automata, lattice-gas CA, we study a model of swarming which displays qualitatively different behaviours under certain experimental conditions. We discussed these observations by relating them to the potential links with certain attributes of the model [48] .

We studied a phase transition that occurs in the Greenberg-Hastings CA reaction-diffusion [5] .

The density classification problem was taken as a typical framework for studying how decentralised computations can be carried out with simple cells. Although it is known that this problem can not be solved perfectly, we showed that using randomness provides a solution with an arbitrarily high success rate [17] . We also studied how to extend this result to the infinite-space case citerefPapier ???.

We studied the behaviour of the amoebae aggregation model [33] and applied the aggregation scheme on a robotic case (ALICE robots and Khepera III with Romea interactive table).

Ant algorithms for multi-agent patrolling

Participants : Olivier Simonin, François Charpillet, Olivier Buffet, Arnaud Glad.

We proposed in 2007 an ant algorithm, called EVAP, to deal with multi-agent patrolling, which is based on the marking and the evaporation of a digital pheromone. During the simulations carried out to measure the performances of EVAP, we identified that the system can self-organize towards stationary cycles (a periodic attractor). These cycles correspond to an Hamiltonian or quasi-Hamiltonian covering of the environment, which is an optimal or quasi-optimal solution to the multi-agent patrolling problem. We then established the mathematical proof that the system can stabilize only in cycles, one per agent, having the same length (cf. publication in ECAI'2008). Moreover, we introduced new heuristics in the agent behavior that improve dramatically the time for convergence, and we proved that under deterministic hypotheses the system always converges to stable cycles (these results have been published in SASO 2009, AAMAS'10). Results of 2011 are :

• Defense of Arnaud Glad's PhD. thesis (November 15th) synthesising theoretical and experimental studies of the EVAP algorithm. The writing of a journal article is also in progress.

• EVAP has been adapted to continuous space in the context of the SUSIE project, which consider the surveillance of an area with a set of autonomous aerial robots.

Multi-robot systems : swarm intelligence, cooperation, navigation

Multi-robot exploration and mapping

Participants : Olivier Simonin, François Charpillet, Antoine Bautin.

In the context of the ANR Cartomatic project, introduced in Sec. 8.2.3 , we study multi-agent models for multi-robot deployment and mapping. This work is in line with the PhD thesis of Antoine Bautin, started in November 2009. New results of 2011 are

• A new frontier assignation algorithm for multi-robot exploration has been proposed. It relies on counting the number of robots towards a frontier rather than considering only distances between robots and frontiers. We measured on benchmarks that the approach outperforms the two classical algorithms closest frontier and Greedy approach. Results are prensented in [37] , [43] and are submitted to ICRA'2012.

• We implemented and experimented the approach with autonomous mobile robots in the context of the ANR Carotte challenge (June 2011, Bourges). Our team “cartomatic” obtained one of the best map of the contest, while deploying several robots.

New experimental device: the Interactive Table

Participants : Olivier Simonin, François Charpillet, Nicolas Beaufort.

Olivier Rochel (INRIA research engineer, SED Nancy) is an external collaborator.

During 2010 we developed with the Nancy INRIA SED(Service d'Expérimentation et Développement) (Olivier Rochel) a new experimental device dedicated to swarm robotics study. It is composed of two independent components : an interactive table able to display and to compute any active environment and a set of autonomous mobile robots able to read and write information on the environment.

Studies using the Table in 2011 are :

• We revisited the Drogoul & Ferber Foraging model, inspired by ants and also called “robot dockers” as the agents exchange the transported resources when they meet. From this simulated model we examined how it can be implemented with real mobile robots on an interactive environment, by considering that robots drop pheromones as ants. We defined a model extending the docker model with the robots on the Table, and studied its robustness to perception failure/mistakes. This work, done with Thomas Huraux (Master 2 Recherche internship), has been published in ICTAI 2011 Int. Conference [31] .

• Several students (from Science Cog. Nancy 2 Master) implemented and explored pheromone-based foraging behaviors and flocking-based navigation models (supervized by François Charpillet and Christine Bourjot).

Local control based platooning

Participants : Alexis Scheuer, Olivier Simonin, François Charpillet, Jano Yazbeck.

We consider decentralised control methods to operate autonomous vehicles at close spacings to form a platoon. We study models inspired by the flocking approach, where each vehicle computes its control from its local perceptions. We investigate different decentralised models in order to provide robust and scalable solutions. Open questions concern collision avoidance, stability and multi-platoon navigation.

Coupling lateral and longitudinal controls. A first work [67] focused on longitudinal control, which aims at computing velocities to avoid collision when all the vehicles are moving along a fixed path. When vehicles move in a two dimensional space, a lateral controller is needed to steer the vehicles. While lateral and longitudinal controls can be considered separately, the longitudinal control should be done after the lateral control: while turning, a higher inter-vehicle distance is needed to avoid collisions.

An innovative approach to improve the quality of lateral control has been proposed during Jano Yazbeck's internship at LORIA (03/10–07/10), entitled “Decentralised local approach for lateral control of platoons” and supervised by A. Scheuer and O. Simonin. This allows to reduce the distance between each vehicle's path and the path of the previous vehicle, by using only embedded sensors such as a laser rangefinder. It relies on memorizing and computing in real time the previous vehicle relative trajectory. This work has been published in 2011 IEEE-RSJ International Conference on Intelligent Robots and Systems (IROS'2011) [34] .

Finding an efficient lateral control. To obtain an even better lateral control, and to drive each vehicle exactly in the trace of the previous one, we are developing a more efficient lateral control law. This law is defined in order to reduce exponentially the tracking error (which is more or less the distance between each vehicle's path and the path of the previous vehicle). Once again, as for the longitudinal control [67] , the formula of the control law is obtained through the proof of its property: necessary conditions are simplified in order to get the final result.

Adaptation of autonomous vehicle traffic to perturbations

Participants : Mohamed Tlig, Olivier Simonin, Olivier Buffet.

In the context of the european InTraDE project, one problem is to handle the displacements of numerous IAVs(Intelligent Autonomous Vehicle) in a seaport. Here we assume a supervisor planning the routes of the vehicles in the port. However, in such a large and complex system, different unexpected events can arise and damage the traffic : failure of a vehicle, human mistake while driving, obstacle on roads, local re-planning, and so on.

We started focusing on a first important sub-problem of space resource sharing among multiple agents: how to ensure the crossing of two opposed flows of vehicles on a road when one of the two paths is blocked by an obstacle, e.g., a disabled vehicle. To overcome this problem, blocked vehicles have to coordinate with vehicles of the other side to share the road and manage delays. The objective is to improve traffic flow and reduce the emergence of traffic jam.

Solving this problem with reactive coordination methods is a major challenge of the PhD thesis of Mohamed Tlig (started in December 2010).

• We started by formalizing the problem and the possible actions of agents (vehicles) following a STRIPS formalism. We adapted this model dedicated to planning to the description of local rules in reactive coordination.

• We then defined and studied in simulation two decision rules that produce two different strategies: the first one alternates between two vehicles from each side of the road, and the second one gives priority to the vehicle with the highest delay. We are preparing a publication of these first results.

Ambient intelligence and Actimetry

Robotics and spatial computing : the iTiles - intelligent tiles - model

Participants : Olivier Simonin, François Charpillet, Lionel Havet.

Olivier Rochel (INRIA research engineer, SED Nancy) is an external collaborator.

In the context of intelligent home and assistant robots, we explore the definition and use of an active floor based on a cellular network approach. We aim at exploring spatial calculus models when considering physical cells augmented with sensors where robots and humans can evolve. Since 2009, we study a model consisting in paving the floor with interconnected tiles. Each tile can communicate with its neighbors and can sense the presence of a robot or a human. A first Tile model has been defined and evaluated using a tiles emulator and real mobiles robots (Kheperas III), which validated the interest of the approach. See CAR'2010 publication [68] .

In 2011 we designed with the help of INRIA Grenoble SED a prototype of 9 physical tiles embedding a WSN node able of computation and communication with other tiles. From this experimental device we explored several questions :

• How to follow a person walking on such a discrete and sensitive floor ? We proposed a set of distributed algorithms allowing tiles to track a person or a robot (cooperation between neighboring tiles).

• How to make communications between a robot and the tile(s) it occupies ? We developped a set of functions using the wifi communication of the tiles and the robot.

• From the work mentionned in the previous items we propose the definition of a new tile network based on the SensLab technology (wire connections between tiles and wireless communications between tiles and robots/humans). This prototype have been ordered and will be installed in the begining of 2012.

Bayesian 3D Human Motion Capture Using Factored Particle Filtering

Participants : Abdallah Dib, Cédric Rose, Amandine Dubois, François Charpillet.

The gait deterioration of elderly people is an important factor in loss of autonomy and it increases the risk of falls. In order to evaluate this risk the MAIA team has been developing since 2003 a markerless human motion capture system that estimates the 3D positions of the body joints over time. The system uses a dynamic Bayesian network and a factored particle filtering algorithm. This year, we have evaluated the impact of using different observation functions for the Bayesian state estimation: chamfer distance, a pixel intersection and finally a pseudo-observation of the subject direction calculated from the previous output of the system. We also compared two methods for the factored generation of the particles. The first one uses a deterministic interval exploration strategy whereas the second one is based on an adaptive diffusion. The capacity of the system to recover after occlusion by obstacles was tested on simulated movements in a virtual scene [57] .

An other achievement of the year has been the assessment of the accuracy and precision of this system, especially for measuring the step length of a walking human. This has been realized by Amandine Dubois during her research master [58] . An experiment with young subjects has been designed and realized. Measures of the markerless motion capture system were then compared with real values. These values were obtained through the footprints left by the subjects. Ink swabs placed at the front and rear of the shoes of each subject make it possible to mark a paper strip positioned on the ground. A statistical analysis of the results has been done by Amandine. Thus we were able to determine if the real and measured lengths were significantly different or not.

Automatic Evaluation of Vascular Access in Hemodialysis Patients

Participants : Cédric Rose, François Charpillet.

The vascular access that allows to perform the extra-corporeal circulation, is usually a vein of the arm that has been enlarged by a surgical creation of a fistula. The prevention of complications such as stenosis or thrombosis of the vascular access is a key issue in hemodialysis treatment. Many dialysis machines measure ionic dialysance by conductivity measures on the dialysate fluid. Ionic dialysance is an indicator of small molecules transfers through the dialysis membrane. Previous works have shown that the follow-up of the dialysance and the pressures along the extra corporeal circuit can help to detect at an early stage a potential complication on the vascular access. The difficulty of automating the follow-up is the large variability of the measures and the need to detect tendencies. Dynamic Bayesian networks (DBN) allow to formalize expert knowledge as a graphical stochastic model adapted to reasoning under uncertainty. In a DBN the state of the patient and the measurements are represented by interconnected temporal random variables. The relations between those variables are described using probability distributions. The proposed approach [64] is based on a supervised learning of a DBN for classifying the dialysis sessions according to a risk score describing the medical situation (0: no risk, 1: mild risk, 2: severe risk). The training of the system was performed using a dataset labeled by a medical expert. The evaluation of the results was done by performing a double-blind analysis of real data. The result was an 85% agreement rate between the human expert and the automated analysis. The purpose of the system is to assist the human expert by reporting abnormalities. The results show that a score 2 reported by the human is rarely missed by the automated analysis (only 1 case) whereas the opposite is more frequent (8 cases). The final decision to further investigate a case is taken by the human expert.