Document multi-objective LLaMEA options and example usage

docs/Quickstart.rst

@@ -31,8 +31,9 @@ Quick Start
 Examples
 --------
 
-Below are three example scripts demonstrating LLaMEA in action for black-box
-optimization with a BBOB (24 noiseless) function suite, and one Automated Machine Learning use-case.   
+Below are four example scripts demonstrating LLaMEA in action for black-box
+optimization with a BBOB (24 noiseless) function suite, a multi-objective
+optimization workflow, and one Automated Machine Learning use-case.
 One of the black-box optimization scripts (`example.py`) runs basic LLaMEA, while the other (`example_HPO.py`) incorporates
 a **hyper-parameter optimization** pipeline—known as **LLaMEA-HPO**—that employs
 SMAC to tune the algorithm’s parameters in the loop.
@@ -124,4 +125,35 @@ In this example, a basic classification task on the breast-cancer dataset from s
 
 .. note::
    Adjust the model name (`ai_model`) or API key as needed in the script.
-   You can easily change the dataset, task and evaluation function to fit your needs.
+   You can easily change the dataset, task and evaluation function to fit your needs.
+
+
+Running ``multi_objective.py``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+**``multi_objective.py``** demonstrates Pareto-based optimization with
+LLaMEA on a synthetic Travelling Salesman Problem variant that optimizes
+two conflicting objectives:
+
+- **Distance**: total route length.
+- **Fuel**: route cost with load-dependent fuel consumption.
+
+The example highlights:
+
+- Returning a :class:`~llamea.multi_objective_fitness.Fitness` object from
+  the evaluator.
+- Enabling ``multi_objective=True`` in :class:`~llamea.llamea.LLaMEA`.
+- Passing ``multi_objective_keys=["Distance", "Fuel"]`` so objective values
+  are tracked consistently.
+- Receiving a :class:`~llamea.pareto_archive.ParetoArchive` and extracting the
+  final non-dominated set.
+
+How to run:
+
+.. code-block:: bash
+
+   python multi_objective.py
+
+.. note::
+   The script defaults to an Ollama model (``gemma3:12b``). Update the LLM
+   backend and credentials to match your local setup.

docs/llamea.rst

@@ -15,6 +15,9 @@ Recent features include:
   instead of entire source files from the LLM. This is more token efficient for large code bases.
 * **Population evaluation** – with ``evaluate_population=True`` the evaluation
   function ``f`` operates on lists of solutions, allowing batch evaluations.
+* **Multi-objective mode** – set ``multi_objective=True`` and provide
+  ``multi_objective_keys=[...]`` to optimize multiple objectives and maintain a
+  Pareto archive instead of a single best solution.
 * **Warm start** -With every iteration, **LLaMEA** archives its latest run in
     `<experiment_log_directory>/llamea_config.pkl`. The framework provides
     **``warm_start`` class methods** that allow you to resume from a previously
@@ -52,6 +55,8 @@ The most important keyword arguments of :class:`LLaMEA` are summarised below.
      - Prompt engineering controls.
    * - ``mutation_prompts`` / ``adaptive_mutation`` / ``adaptive_prompt``
      - Mutation and prompt adaptation settings.
+   * - ``multi_objective`` / ``multi_objective_keys``
+     - Enable Pareto-based optimization and define objective names.
    * - ``budget`` / ``eval_timeout`` / ``max_workers`` / ``parallel_backend``
      - Runtime and parallelisation controls.
    * - ``log`` / ``experiment_name``

examples/multi_objective.py

@@ -1,3 +1,12 @@
+"""Multi-objective LLaMEA example on a synthetic TSP variant.
+
+This script shows how to:
+1. Evaluate generated code against two objectives (Distance and Fuel).
+2. Return objective values using ``Fitness``.
+3. Run LLaMEA with ``multi_objective=True`` and objective keys.
+4. Read final non-dominated solutions from ``ParetoArchive``.
+"""
+
 import os
 import random
 from typing import Optional
@@ -26,6 +35,7 @@ def __repr__(self):
 
 
 def generate_tsp_test(seed: Optional[int] = None, size: int = 10):
+    """Generate a depot and customer set for the synthetic TSP task."""
     if seed is not None:
         random.seed(seed)
     depot = Location(0, 50, 50, 0)
@@ -46,6 +56,12 @@ def generate_tsp_test(seed: Optional[int] = None, size: int = 10):
     referable_dict[customer.id] = customer
 
 def evaluate(solution: Solution, explogger: Optional[ExperimentLogger] = None):
+    """Evaluate generated solver code on a two-objective TSP benchmark.
+
+    The generated class must return a permutation of customer ids. The evaluator
+    validates the route, computes total travel distance and load-dependent fuel
+    usage, then stores a ``Fitness`` object with both objectives.
+    """
     code = solution.code
 
     global_ns, issues = prepare_namespace(
@@ -180,7 +196,8 @@ def __call__():
         return customer_ids
 """
 
-    llamea_inst = LLaMEA(f=evaluate, 
+    # Multi-objective mode returns a Pareto archive instead of a single winner.
+    llamea_inst = LLaMEA(f=evaluate,
            llm=llm,
            multi_objective=True,
            max_workers=3,
@@ -192,10 +209,11 @@ def __call__():
            example_prompt=example_prompt,
             experiment_name="MOO-TSP",
             minimization=True,
-            budget=27
+           budget=27
            )
 
     solutions = llamea_inst.run()
+    # Keep only the final non-dominated set for reporting/inspection.
     if isinstance(solutions, ParetoArchive):
         solutions = solutions.get_best()
 
@@ -206,4 +224,4 @@ def __call__():
         print(solutions.description)
         print(solutions.code)
         print(solutions.fitness)
-        print("------------------------------------------------------------------------------------------------------------------------")
+        print("------------------------------------------------------------------------------------------------------------------------")

llamea/llamea.py

@@ -110,6 +110,14 @@ def __init__(
             task_prompt (str): A prompt describing the task for the language model to generate optimization algorithms.
             example_prompt (str): An example prompt to guide the language model in generating code (or None for default).
             output_format_prompt (str): A prompt that specifies the output format of the language model's response.
+            multi_objective (bool): Enable multi-objective optimization mode.
+                When set to ``True``, the evaluation function should assign a
+                :class:`~llamea.multi_objective_fitness.Fitness` object via
+                :meth:`~llamea.solution.Solution.set_scores`.
+            multi_objective_keys (list[str]): Ordered objective names used by
+                the multi-objective pipeline (e.g. ``["Distance", "Fuel"]``).
+                Each key must be present in every returned
+                :class:`~llamea.multi_objective_fitness.Fitness` object.
             experiment_name (str): The name of the experiment for logging purposes.
             elitism (bool): Flag to decide if elitism should be used in the evolutionary process.
             HPO (bool): Flag to decide if hyper-parameter optimization is part of the evaluation function.