Evolutionary Systems

Evolutionary computation is based on the idea that intelligence can be discovered by simulating the process of natural selection: variation, selection, and inheritance.

Key Concepts

Genetic Algorithms (GA)

Representing potential solutions as "chromosomes" (often bitstrings) and evolving them through crossover and mutation.

Genetic Programming (GP)

Evolving computer programs rather than just data. Programs are often represented as trees or graphs.

Neuroevolution

Evolving the structure and weights of neural networks.

• NEAT (NeuroEvolution of Augmenting Topologies): A popular algorithm that starts with simple networks and evolves complexity over time.

Evolutionary Computation in AGI

Open-ended Evolution

The goal of creating systems that continuously generate novel and increasingly complex behaviors, rather than just solving a fixed task.

MOSES (Meta-Optimizing Semantic Evolutionary Search)

A component of OpenCog/Hyperon that evolves program fragments in a representation-agnostic way. It uses local search to find high-performing programs and creates new ones based on identified patterns.

Strengths

• Non-Differentiable Search: Can optimize systems where gradients aren't available (e.g., discrete logic, hardware design).
• Exploration: Genetic diversity helps prevent the system from getting stuck in local optima.
• Architecture Search: Can discover novel neural architectures (NAS) that humans wouldn't think of.

Weaknesses

• Computational Cost: Often requires many generations and evaluations, which can be expensive.
• Scaling: Maintaining diversity in very large search spaces is challenging.

---

Next: Probabilistic Architectures