ICLP 2005 Doctoral Consortium Concurrent Methodologies for Global Optimization

Introduction. My research interests are mainly focused on concurrent approaches to global optimization problems. Optimization tasks have two main conflicting features: they are both very difficult and central to a lot of the applications computer science faces daily. The problems I’m most interested in stem out from biology, protein folding being the principal one1. The protein folding, or protein structure prediction problem, is the task of identifying the characteristic (native) spatial configuration of a protein (a polymeric chain built up from 20 different aminoacids), given the sequence of aminoacids composing it. This problem can be modeled as the search of the configuration of minimum free energy, and even very coarse abstractions of it are known to be NPhard. Needless to say, it is far from being solved in a satisfactory way, even if there is a huge amount of research on it, due to its enormous importance in biological and pharmaceutical research. The most striking feature, however, is not its mathematical difficulty, but the fact that Nature is able to fold correctly a protein in a very short time (milliseconds), “exploring” a very small portion of the search space, even if the main forces behind the process (i.e. protein-solvent interaction) have a “stochastic” nature (in the sense of statistical mechanics).2 In addition, the native configuration is reached simply by the interaction among the atoms constituting the protein. In some sense, it’s the concurrent interaction between these “simple” agents, that obey “simple” rules (the laws of physics), that determines both the protein’s native structure and the dynamics of the process for reaching it. Therefore, one of the key ingredient that enables Nature to be so efficient could be this total concurrency itself. Concurrent simulation of Protein Folding. The previous reflections are the starting point of the attempts we made so far in modelling the protein structure prediction problem in a concurrent setting. In [1], we associated an independent process to each aminoacid, modeled here in a simplified way as a single centre of interaction. These agents interact by exchanging information about their spatial position, using this knowledge to move in the space, trying to reach the configuration of minimum free energy. The moving strategy used is a Monte-Carlo one: moves lowering the energy are always performed, while moves rising the energy are executed with a certain probability depending on the difference of potential. The whole simulation was written in SICStus-LINDA. The results were encouraging, even if the coarseness of the energy model used and the slowness of communications between LINDA processes forbade to obtain decent predictions for whole proteins. To tackle these problems, in [2] we embedded the previous framework in a multi-agent scheme designed for optimization tasks, implementing the whole system both in SICStus-LINDA and in C (in a multi-threading version). In this new model we adopted a different potential, still representing the aminoacids as a single center of interaction, and identifying them with a concurrent process. In addition, we introduced some higher level agents, which have the task of both coordinating the exploration of the state space performed by the aminoacid agents and introducing some form of cooperation among them. This cooperative action helps the aminoacids to form some local patterns, like helices and sheets, that are very common in proteins. The enhancements introduced here improve a lot the