Special Cases in Distributed Execution

While most metaheuristics in MetaGen follow a standard distributed execution model using Ray, some algorithms require additional considerations due to their specific computational structures. This section details two specialized cases:

Memetic Algorithm: Allows fine-grained control over distribution levels.
Distributed CVOA: Implements a complex epidemic-based optimization model with strain-level parallelism.

Memetic Algorithm and Distribution Levels

The Memetic Algorithm extends standard evolutionary strategies by incorporating local search to refine solutions within each generation. In the distributed context, it introduces the concept of distribution levels, which control how computational tasks are assigned across available resources.

Enabling Distribution in Memetic

When initializing the Memetic algorithm, users can specify the distributed parameter and an additional ``distribution_level``:

from metagen.metaheuristics.mm import Memetic

memetic = Memetic(domain=my_domain,
                  fitness_function=my_fitness_function,
                  population_size=100,
                  mutation_rate=0.1,
                  neighbor_population_size=10,
                  alteration_limit=1.0,
                  distributed=True,
                  distribution_level=1)

best_solution = memetic.run()

Understanding Distribution Levels

The distribution_level parameter defines how computation is parallelized:

Level 0 (default) – No distributed execution; all computations run sequentially.
Level 1 – Distributes local search operations (e.g., mutation and refinement of solutions) across available CPU cores.
Level 2+ – Extends distribution to additional genetic operations, further increasing parallelism.

This flexibility allows users to balance overhead vs. performance gain, ensuring that distribution is beneficial rather than introducing unnecessary communication costs.

When to Use Distribution in Memetic

Recommended for large populations where local search is computationally expensive.
Not ideal for small-scale optimizations where overhead may outweigh speed improvements.

—

Distributed CVOA: Multi-Strain Parallelization

The Coronavirus Optimization Algorithm (CVOA) in MetaGen follows a unique epidemic-inspired optimization model. Its distributed version extends this by allowing multi-strain execution, where different “viral strains” explore the search space in parallel.

How CVOA Executes in a Distributed Environment

Unlike traditional metaheuristics, Distributed CVOA manages a global state using Ray actors. Each strain operates independently while interacting with a shared environment.

Launching Distributed CVOA

Users must initialize Ray and execute multiple strains concurrently using the distributed_cvoa_launcher function:

from metagen.metaheuristics.cvoa import distributed_cvoa_launcher

best_solution = distributed_cvoa_launcher(strains=my_strains,
                                          domain=my_domain,
                                          fitness_function=my_fitness_function,
                                          update_isolated=True,
                                          log_dir="logs/DCVOA")

Key Features of Distributed CVOA

Strain-Based Parallelism Each strain represents an independent execution of CVOA, enabling multiple parallel searches.
Remote Pandemic State Management A global state tracks recovered, dead, and infected individuals across all strains.
Ray-Based Worker Management Tasks such as infection spreading, mutation, and selection are dynamically assigned to available computing resources.

Example Log Output

During execution, logs provide real-time insight into the epidemic progression:

[Strain 1] Best individual fitness: 0.0291
[Strain 2] Best individual fitness: 0.0312
Pandemic Report: Recovered = 152, Deaths = 21, Best Individual = Solution(0.0291)
Execution time: 00:02:14

When to Use Distributed CVOA

For large-scale optimizations requiring complex exploration strategies.
When multiple independent searches need to run in parallel.

—

Comparison of Distributed Execution Approaches

Metaheuristic	Distribution Strategy	Recommended Use Cases
Memetic	Selective distribution based on `distribution_level`	Large populations, local search is computationally expensive
CVOA	Strain-based parallelization with a global pandemic state	Large-scale optimization with epidemic-inspired search models

By leveraging these specialized distributed execution methods, MetaGen allows users to optimize metaheuristics efficiently, ensuring that distribution aligns with algorithmic needs.