Prolog planner generation prompting guide
a year ago
- #Planning
- #LLM
- #Prolog
- LLMs excel in language tasks but struggle with planning outside their training data.
- Combining LLMs with Prolog leverages Prolog's combinatorial power for planning tasks.
- LLMs are better suited for translating natural language into Prolog rather than direct planning.
- Prolog's natural language processing roots make it a good target for LLM-generated code.
- Practitioners prefer Prolog over domain-specific languages due to its logical underpinnings.
- Chain-of-thought (CoT) approaches improve LLM reasoning but require extensive training data.
- LLMs can generate Prolog code for state, action, and check predicates but not the main solver.
- WARPLAN, an early Prolog-based planner, influenced modern Prolog planning systems.
- Using asserta/1, assertz/1, and retract/1 in Prolog simplifies state changes for LLM-generated code.
- Realistic logistics problems require additional components like business rules and user interfaces.