Reinforced Linear Genetic Programming

AI summary

Overview

Research area: Evolutionary computation, genetic programming, reinforcement learning, and automated program synthesis.
Technical level: Intermediate.
Scope: This thesis proposes Reinforced Linear Genetic Programming (RLGP), which combines Q-Learning with linear genetic programming to automate register-action assignments, and introduces a memory-safe Rust framework called linear-gp for experimentation.

What This Paper Is About

Linear Genetic Programming (LGP) evolves programs represented as linear instruction sequences, but it typically requires humans to explicitly map registers to actions. This thesis addresses that manual bottleneck by using Q-Learning to learn optimal register-action assignments automatically. The goal is to make LGP more autonomous and to provide a Rust-based framework for further research.

Key Contributions

1. Proposes Reinforced Linear Genetic Programming (RLGP), a novel approach that applies Q-Learning on top of LGP.
2. Learns optimal register-action assignments, reducing or removing the need for humans to map registers to actions manually.
3. Introduces "linear-gp," a new framework written in memory-safe Rust for extensive experimentation.
4. Establishes a foundation for future work combining reinforcement learning with linear genetic programming.

Main Findings

• Automated register-action learning: RLGP uses Q-Learning to discover optimal register-action assignments instead of relying on human-designed mappings.
• Limitation addressed: The approach directly targets a known limitation of traditional LGP: the need for explicit human mapping of registers to actions.
• Rust framework introduced: The "linear-gp" framework provides a memory-safe implementation intended to support wide-ranging experiments.
• Foundation for reinforced LGP: The work opens a path for further research into reinforcement learning-enhanced genetic programming.
• No quantitative benchmarks in abstract: The abstract states the proposed approach and framework but does not report performance metrics or comparisons.

Methodology in Plain English

In LGP, programs are lists of instructions that read from and write to registers. Normally, a human decides which register corresponds to which action. This thesis treats those assignments as a learning problem. A Q-Learning agent tries different register-action mappings, observes how well the resulting LGP programs perform, and updates its choices over time. The LGP process evolves the instruction sequences, while Q-Learning tunes the mapping between registers and actions. The entire system is implemented in Rust to ensure memory safety and enable experimentation.

Why This Matters

Impact on research: This work bridges evolutionary computation and reinforcement learning, offering a way to reduce manual design in genetic programming and providing open tooling for reproducible experiments.

Real-world applications:

• Automated discovery of mathematical models and symbolic regression.
• Program synthesis and automatic code generation.
• Control policy learning for robotics and simulated agents.
• Compiler or virtual machine optimization where register and instruction mapping affects performance.

Industry relevance: Companies using automated machine learning, simulation, embedded control, or code synthesis could benefit from reducing manual register-action mapping. The Rust framework may appeal to production environments that require memory safety and performance.

Future Directions

• Benchmark RLGP against standard LGP and other reinforcement learning plus genetic programming hybrids.
• Test alternative reinforcement learning algorithms beyond Q-Learning for register-action assignment.
• Scale RLGP to larger register sets, more complex instruction sets, and diverse problem domains.
• Extend the "linear-gp" Rust framework with visualization, parallelization, and integration with existing GP or RL ecosystems.
• Analyze computational cost, convergence behavior, and robustness of the learned register-action mappings.

Target Audience

Graduate students and researchers in evolutionary computation, genetic programming, reinforcement learning, and automated program synthesis. Developers interested in Rust-based AI frameworks. Practitioners in AutoML, automated software engineering, and control systems who want to understand how reinforcement learning can reduce manual design in program evolution.

Authors’ abstract

Linear Genetic Programming (LGP) is a powerful technique that allows for a variety of problems to be solved using a linear representation of programs. However, there still exists some limitations to the technique, such as the need for humans to explicitly map registers to actions. This thesis proposes a novel approach that uses Q-Learning on top of LGP, Reinforced Linear Genetic Programming (RLGP) to learn the optimal register-action assignments. In doing so, we introduce a new framework "linear-gp" written in memory-safe Rust that allows for extensive experimentation for future works.