MOSES

MOSES (Meta-Optimizing Semantic Evolutionary Search) is Hyperon's primary engine for program synthesis. Instead of evolving raw bitstrings, MOSES evolves short programs represented in a specialized representation.

How MOSES Works

MOSES is designed to find programs that satisfy a specific objective function (fitness).

1. Representation

MOSES uses a hierarchical representation that is easier to evolve than raw MeTTa or Python code. It avoids "junk" or "bloat" that often plagues standard Genetic Programming.

2. Semantic Search

Unlike traditional GAs that rely on random mutation, MOSES uses a more "intelligent" search process:

• It maintains a population of program templates.
• It uses local optimization to "tune" the parameters of these templates.
• It identifies high-performing "building blocks" and combines them.

3. Competence Monitoring

The algorithm monitors its own performance and adjusts its search strategy based on the complexity of the task and the quality of the candidates found so far.

Role in Hyperon

MOSES is used for:

• Feature Discovery: Finding novel patterns in large datasets.
• Automatic Program Improvement: Evolving more efficient versions of existing algorithms.
• Goal-Directed Synthesis: Creating a plan or behavior to achieve a specific goal.

MOSES vs. Deep Learning

• Deep Learning: Powerful at finding patterns in continuous spaces with gradients.
• MOSES: Powerful at finding patterns in discrete, structured spaces (like code) where gradients are not available.

In a full AGI system, MOSES might be used to evolve the "high-level logic" or "strategies" that control a lower-level neural network.

---

Next: The Pattern Matcher