A Genetic Source-Code Program-Synthesizer for Real-Time Co-Evolution with Humans: Gene-Level Geometric-Push Program-Synthesis

Can genetic operators be used to produce human-centric source-code documents for human-interactive and collaborative programming practice?

Addressing this question, a novel automated source-code generation system is created which synthesizes, using genetic operators, object-oriented compilable files, which better conforms to human coding conventions than current benchmark genetic algorithms. This project:

• Provides a new form of AI for automatic source-code generation.
• Implements novel use cases for automatic source-code generation in practical collaborative applications.
• Benchmarks the new algorithm against a standard program synthesis benchmark suite.
• Assesses a genetic realignment algorithm in the context of crossover for elitest fitness.
• Explores novel approaches to interaction methods with a genetic auto-coder for multiple simultaneous users.

As of the beginning of this project, there were no existing implementations of sourcecode generators designed for human legibility while synthesizing for arbitrary language environments: automated program synthesis algorithms derived from genetic operators traditionally operate on specialist languages, with bytecode or machine code solutions or suffered significantly with comprehension against human written code.

The novel algorithm introduced in this thesis interjects into current discourse with a heuristic approach which can operate with relative language agnosticism, while retaining common coding conventions, indentation, arbitrary code length, looping operators, multiple function definitions and function calls.

Using this new algorithm, this thesis explores three diverse experimental phases to analyse potential use-cases, from a qualitative perspective:

An automatic programmer for coding problems: as a tool to support human software developers, for simple programming tasks.

A co-creative environment with Artificial Life: demonstrating the automatic programmer’s ability to create evolvable behaviour controllers which adapt to survive under various environmental pressures and to co-evolve with human interactors.

A public facing music generator: as a collaborative medium for live performance environments, providing human-guided fitness training for live evolution of audio.

The thesis concludes that the algorithm is successful in automatic coding for implicit fitness or regression environments, with several limitations relating to the fitness function used and the size of the search space. These limitations are established and persist across genetic methods for automatic coding. Regardless of the limitations, this approach demonstrates valuable use cases in artistic mediums.