'Cheap grid': Leveraging system failure using stochastic computation

Traditionally, network and computation failure on a heterogeneous network are viewed as an unfortunate obstacle to reliable, efficient computation. We propose that such noise can be incorporated into the algorithm design as part of the necessary source of randomness used in stochastic computation. This paradigm incorporates network and computation failure at a high level in the solution-discovery algorithm, rather than attempting to hide and suppress all such noise at the lowest possible levels in the computation tool. This idea enables the creation of a network solution system with extremely small amounts of global state. This lack of required system state allows for heightened degrees of scalability in the computation engine, and fewer resources are consumed by system management. Algorithms with a stochastic component are easily adapted to this system; various types of evolutionary computation are particularly well adapted to this hybrid paradigm. A specific example using a modified steady-state genetic algorithm is provided to explore the functionality of the resulting composite system. The developed architecture is used to calculate the solution to a number of problems, in each case converging on a solution measurably faster than that of a fully ''fault-tolerant'' scheme, thereby resulting in lower overhead and faster execution time.