A major obstacle in the simulation was the time it would take for agents to generate a random production based on the evolving probabilities of their articulatory space. The idea was that for any production, a continuous value for place of articulation (the degree would be kept constant for now) would be generated and then the acoustic output would be computed. It takes at least some 2 seconds to calculate each vowel. Multiplying that by the number of agents anf the thousand/millions of turns, it would take a lot of time to run a single simulation. The solution is to make this production digital and store previously all the possible results. It will be much faster to store some 100 or 200 places of articulations (vowels only) linearly spaced. The computation time would be replaced for memory access time. With at least 100 linearly spaced points, the continuity would be present and the non-linearity of the articulatory-acoustic relation would remain intact.
With this change, it will be possible to run simulations much faster than what I estimated before and also to test most of the parameters ranges which produce stable conditions. The goal now is to have a simulation using a real vocal tract for the Abralin conference next March.
