AMYBIA
Aggregating MYriads of Biologically-Inspired Agents

Milestones of the project

• Determining observable functions: Starting from the prototype, our aim will be to find the best observation functions, i.e., the most adequate parameters that will allow to quantify the global dynamics of the system. These observation functions have to be built according to a trade-off between the precision with which they account for the system evolution and low computational cost. A similar study may be found in [BQ06] for recurrent neural networks or in [B03] for cell-cell communication in mammals..
• Analytical and simulation studies of the system statistical features: To understand and exploit the model properties, we will have to estimate several of its statistical properties such as the characteristic time to agent aggregation, the probability distribution and spatial distribution of the aggregates, the universality class of the phase transition or the size-dependence of the system.
• These studies will be carried out under four different environment conditions:
  • base conditions (no obstacles, no defects, regular neighbourhood);
  • using various types of neighbourhood topologies (regular, small-world, scale-free);
  • in the presence of obstacles, with variable geometries and sizes;
  • with different types or failures/defects (environment, agents, rules).
• Using FPGA implementation of the model as a sanity check of its feasibility. This part of the work addresses different hardware parallelism issues: compact design of mixed cellular automata and agents, implementation of concurrent transition rules, simulation of moving agents, local and area-saving handling of stochastic laws, simulation of asynchronism, etc.

  These developments will direct adaptations of the theoretical model and of its statistical study, while offering large scale simulation capabilities. The computational and topological flexibility of FPGAs will finally help us evaluate our approach in terms of decentralized low-cost control threads.

• Comparing the performance of our system with other existing methods (ants algorithms, reaction-diffusion processes) as far as fault/defect tolerance, behaviour in the presence of obstacles or irregular neighbourhood relationships are concerned.

Expected benefits and broader expectations

• promoting a new way of conceiving complex systems : the theory-hardware codesign
• developing tools for efficient simulations on FPGAs
• proposing an elegant and effective solution to the decentralised gathering problem
• comparing our solution to other solutions such as using virtual ants with diffusion-evaporation schemes

We also expect that besides the planned research activities, this project will allow us to create synergies with potential collaborators. Two collaborative approaches have already been identified:

• Using more elaborate simulation tools: For the study of the system statistical features, an additional approach is to benefit from the properties of the programming language MGS and to take advantage of its aptitudes to handle neighbourhood and spatial relationships. The idea is to improve the performance and expressivity of our model. This could be developed in collaboration with one of the co-developers of MGS, Olivier Michel (Ibisc, Université d'Evry Val d'Essonne, Genopole).
• Implementing the model in robots: As a further check of the model feasibility and robustness in a different context, we might port our model on a robotic system, using sugarcube Alice robots for the agents and computer-assisted video-projection of colour patterns for the environment. This kind of robotic platform has recently been built to port pheromone-based ant systems to robotics by G. Theraulaz's group in Toulouse (CRCA).