Human like programming using SPADE BDI agents and the GPT-3-based Transformer

Abstract

Programming an application requires multiple people with skills and experience in that field. It will also take a lot of time with multiple steps before achieving the final result of an application. Today, developers are assisted by various tools, software, or applications based on Artificial Intelligence (AI) such as OpenAI's ChatGPT. These AI that automatically generates source code helps developers to develop applications much faster. However, although code generators are numerous and very helpful, we are not yet at the stage where we can generate a fully functional application, but just generate pieces of source code. And we don’t know yet how to understand textual descriptions of Software Requirements to generate an application directly. Or where to find data to train an AI capable of generating a functional application from textual descriptions. Therefore, we created a new architecture composed of virtual intelligent agents called SPADE BDI to create virtual developers. The virtual intelligent agents were responsible for keyword extraction, Software Requirements synthesis, and source file creation. Then we used a transformer based on pre-trained GPT-3 for source code generation. This transformer is orchestrated by a virtual intelligent agent. To solve the problem of training data, we collected and created a new dataset called WSBL. The data came from several projects developed with the Laravel Framework over 4 years. The result allowed us to have a functional application directly from a textual description. Each intelligent virtual agent played a role like a developer by analyzing textual of Software Requirements and then generating source code. With a 15% reduction in time to develop an application compared to brute development. Our new architecture allows for processing textual descriptions (Software Requirements) step by step using intelligent virtual agents named SPADE BDI and source code generation is done by a transformer based on pre-trained GPT-3 to have a directly functional application