Adaptive Buffer Tuning for Data Intensive Algebraic Operations in Purpose of Parallel and Distributed Processing

Both pervasive directional graphs and intensive algebraic operations require buffer management for stochastic data processes with constrained computing resources. Algebraic computation states in final stages tend to be readily identified within finite time horizon by sensing very abrupt transitions in system and network state spaces. But in early stages of constraints, these changes are hard to predict and difficult to distinguish from usual state fluctuations. Dynamic buffer allocation and replacement are the major techniques to construct structures for algebraic operations to ensure finite resource assesses. Hence, dynamic buffering function and control optimization are the major primitives to construct utilities for stochastic system processes to ensure converged resource accesses. To provide adaptation to large dimensional states, this research proposes a formal model-free buffer utility framework rooted from reinforcement learning methods and dynamic programming techniques to provide self organization of buffers to exploit parallel based buffer tuning processes. To time and space complexity reduction within the large state spaces, dynamic hidden neurons with incremental tuning is proposed for non-linear value function approximation to derive optimization procedures for optimal algebraic computational policies. For numeric and information evaluation, convergence analysis and error estimation are presented. Finally, a simulation test-bed and tuning results are deliberated.