Faster GPU-based genetic programming using a two-dimensional stack

Abstract

Genetic Programming (GP) is a computationally intensive technique which also has a high degree of natural parallelism. Parallel computing architectures have become commonplace especially with regards Graphics Processing Units (GPU). Hence, versions of GP have been implemented that utilise these highly parallel computing platforms enabling significant gains in the computational speed of GP to be achieved. However, recently a two dimensional stack approach to GP using a multicore CPU also demonstrated considerable performance gains. Indeed, performances equivalent to or exceeding that achieved by a GPU were demonstrated. This paper will demonstrate that a similar two dimensional stack approach can also be applied to a GPU based approach to GP to better exploit the underlying technology. Performance gains are achieved over a standard single dimensional stack approach when utilising a GPU. Overall, a peak computational speed of over 55 billion Genetic Programming Operations per Second are observed, a two fold improvement over the best GPU based single dimensional stack approach from the literature.

GSGP-CUDA — A CUDA framework for Geometric Semantic Genetic Programming

“…This operation is naively parallel, allowing for a direct implementation in GPUs. Such architectures have been exploited in deep learning [7], but have been sparsely used with GP [8][9][10]. Note that the works presented by Chitty [8], and Langdon and Banzhaf [9,10] focus on implementing CUDA and other parallel tools to accelerate the interpretation of syntax trees, since that is the representation most commonly used in GP.…”

A Survey of Genetic Programming and Its Applications

“…A Python-based environment and stack-based language for genetic programming can be found in Ref [93]. [94], [95], [113]  High Performance Computing Parallel Computing, Vector Processing, GPU processing, etc.…”

“…Type of GP Advantages Disadvantages Tree-based (TGP) [77], [100], [103]  Higher-order functions are a powerful addition to the TGP which enables the evolution of programs with greater than constant-time complexity  Closure (having the same data type between operators and terminals), which causes to increase the AC in the multiple data type problems  High AC in the Lawnmower and H-IFF problems Stack-based (SGP) [80], [94], [95], [113], [114], [120]  High performance on symbolic regression problem  Low AC (outperforms the TGP)  Efficient performance in parallel computing  Inefficient performance where long programs (variables) are pushed in the stack on limited resources systems  It can only be implemented on stack support languages. Linear (LGP) [81]  High flexibility (e.g., allows more freedom on the internal representation)  Low AC  Allowing a more efficient evaluation of programs  Higher compiler overhead than the TGP Grammatical Evolution (GEGP) [78], [83], [100], [121], [124]  The flexibility of language choice that it allows (e.g.…”

A MIMD Interpreter for Genetic Programming

“…More recently, General Purpose Graphic Processing Units (GPGPUs) have also been proposed as a computing platform well-suited for parallel GP processes, with a seminal work in [3], and further extensions focused on sub-tree parallelization [9], two-dimensional stacks [10] and quantum-inspired linear GP [11]. Given that GPUs have thousands of computing units, huge speedups can be achieved since each unit executes the interpreter on a different block of data.…”