high performance computing on graphics processing units: hgpu.org

This paper studies an option for offloading some types of AI processing to the Graphics Processing Unit (GPU), by proposing the parallelization of the Batch Least Squares (BLS) method for tuning consequent parameters and the gradient method for tuning input fuzzy sets in a Takagi-Sugeno-Kang Fuzzy Inference System using the Compute Unified Device Architecture (CUDA). A method is proposed to generate the required intermediary matrices using heavy data parallelism. The learning consists of several iterations of BLS for the output values and gradient tuning for the input fuzzy sets. The explanation of the methods is followed by a performance comparison with a typical CPU-only approach and evaluation of the feasibility of using this method in real-time inside of a game.