A search program explores candidate programs to find one that satisfies a specification or objective. Discrete program search complements deep learning rather than replacing it, especially when systems combine learned models with algorithmic components. section 19.4.5
Search programs appear in hybrid systems in two main roles:
- Combine learned modules with discrete algorithmic modules, as in systems that pair neural perception or policy components with search, reasoning, memory, or world models. section 19.4.5
- Use deep learning to make the program-search process itself more efficient. section 19.4.5
Figure 19.15 A learned program relying on both geometric primitives (pattern recognition, intuition) and algorithmic primitives (reasoning, search, memory)
The figure depicts a learned program combining geometric primitives for pattern recognition and intuition with algorithmic primitives for reasoning, search, and memory. section 19.4.5
