RKO - Random-Key Optimizer (Python Parallel Framework)

The Random-Key Optimizer (RKO) is a high-performance, parallel metaheuristic framework designed for combinatorial optimization problems. By mapping discrete, high-dimensional search spaces onto a continuous unit hypercube $[0, 1)^n$, RKO detaches optimization algorithms from problem-specific logic.

RKO orchestrates up to 8 parallel metaheuristic workers using Python's native multiprocessing. Workers cooperate in real time by feeding and pulling elite solutions from a thread-safe, shared Solution Pool, accelerating convergence and improving optimization robustness.

---

💎 Core Features

• 8 Concurrent Metaheuristics: Run BRKGA, Multi-Start, SA, VNS, ILS, LNS, PSO, and GA concurrently.
• Online hyperparameters tuning via Q-learning
• Collaborative Search: Thread-safe SolutionPool sharing elite solutions to pull workers out of local minima.
• Automatic Result Structuring: Automatically groups execution runs, creating a dedicated timestamped folder (results_{instance}_{timestamp}/) containing:
  • results.txt: Scientific summary containing aggregated averages, standard deviations, lists of costs, and computational times for academic reporting.
  • logs.txt: Chronological trace of improvements (NEW BEST), optimized and clean for parser engines.
  • run_x.png: Convergence trajectory plot generated automatically for each trial.

---

📦 Installation

pip install rko

---

⚙️ Solver Configuration

RKO Class Initializer

from rko import RKO

rko_solver = RKO(
    env=my_environment,
    best_possible=None,     # early-stop threshold
    logger="dual",          # logger strategy or custom LogStrategy instance
    log_filepath="./results/results.txt" # Base directory for structured outputs
)

Solve Method Parameters

final_cost, final_solution, time_to_best = rko_solver.solve(
    time_total=30,      # Total time limit in seconds
    brkga=1,            # Number of parallel BRKGA instances
    ms=1,               # Number of parallel Multi-Start instances
    sa=1,               # Number of parallel SA instances
    vns=1,              # Number of parallel VNS instances
    ils=1,              # Number of parallel ILS instances
    lns=1,              # Number of parallel LNS instances
    pso=1,              # Number of parallel PSO instances
    ga=1,               # Number of parallel GA instances
    restart=1.0,        # Fraction of total time for each restart cycle
    runs=10,            # Number of independent trials
    plot=True           # Enable automatic convergence plotting (generates .png charts)
)

---

📊 Extensible Logging Engine

The RKO logging subsystem is built on the Strategy Pattern, allowing developers to customize logging behavior by passing a subclass of LogStrategy directly to the logger parameter in the RKO constructor.

Built-in Logger Strategies

• "dual" (TerminalLogger with filepath): Prints logs to stdout and saves structured records to the dynamic results directory.
• "terminal" (TerminalLogger): Outputs logs exclusively to stdout.
• "file" (FileLogger): Saves logs exclusively to the specified disk path.

Custom Logging Strategy Example

You can inherit from LogStrategy to route logs to databases, dashboards, or external analytical tools:

from rko.LogStrategy import LogStrategy

class DatabaseLogger(LogStrategy):
    def __init__(self, db_connection):
        self.db = db_connection

    def log(self, *args, **kwargs):
        msg = " ".join(str(arg) for arg in args)
        # Custom logic to write to database
        self.db.insert_log(msg)

🛠️ Environment Interface & Validation

To adapt RKO to your problem, subclass RKOEnvAbstract and implement decoder and cost.

,
import numpy as np

class RKOEnv():
    def __init__(self):
        # Required attributes
        self.tam_solution: int = 50           # Dimension of the random key vector
        self.LS_type: str = 'Best'            # Local Search strategy ('Best' or 'First')
        self.dict_best: dict = {}             # Shared best dictionary
        self.instance_name: str = "kp50.txt"  # Name of your problem instance

        # Configurations for Metaheuristics
        self.BRKGA_parameters = {'p': [1000, 2000], 'pe': [0.2, 0.15], 'pm': [0.05], 'rhoe': [0.7]}
        self.SA_parameters = {'SAmax': [100], 'alphaSA': [0.99, 0.9], 'betaMin': [0.01], 'betaMax': [0.05], 'T0': [10000]}
        self.ILS_parameters = {'betaMin': [0.1], 'betaMax': [0.2]}
        self.VNS_parameters = {'kMax': [8,7,6,5], 'betaMin': [0.05]}
        self.LNS_parameters = {'betaMin': [0.1], 'betaMax': [0.3], 'TO': [1000], 'alphaLNS': [0.99, 0.95, 0.9]}
        self.PSO_parameters = {'PSize': [1000], 'c1': [2.05], 'c2': [2.05], 'w': [0.73]}
        self.GA_parameters = {'sizePop': [1000], 'probCros': [0.98], 'probMut': [0.005]}

  
    def decoder(self, keys: np.ndarray):
        """Maps random keys [0, 1) into a problem-specific discrete representation."""
        pass

  
    def cost(self, solution, final_solution: bool = False) -> float:
        """Returns objective cost (RKO minimizes this; return negated value for maximization)."""
        pass

Static and Functional Verification via check_env

RKO provides a built-in diagnostic utility check_env to validate custom environments before runtime. It performs the following verification passes:

1. Attribute & Type Verification: Verifies the presence and types of all core variables (tam_solution, LS_type, dict_best, instance_name) and metaheuristic parameter dictionaries.
2. Functional Dry-Run: Invokes the custom decoder and cost methods using mock random keys to verify execution safety.
3. Mathematical Compliance: Checks that the returned cost is a scalar float and that the decoder successfully processes vectors of dimensions defined by tam_solution.
4. Actionable Diagnostics: If a check fails, the utility prints detailed stack traces and configuration hints to assist the developer in resolving interface mismatches or bugs.

from rko.Environment import check_env

my_env = MyProblemEnv()
if check_env(my_env):
    print("Environment successfully validated.")

---

👥 Maintainers

• Felipe Silvestre Cardoso Roberto - LinkedIn
• João Victor Assaoka Ribeiro - LinkedIn