Multi-Objective Problem Solving With Offspring on Enterprise Clouds

In this paper, we present a distributed implementation of a network based multi-objective evolutionary algorithm, called EMO, by using Offspring. Network based evolutionary algorithms have proven to be effective for multi-objective problem solving. They feature a network of connections between individuals that drives the evolution of the algorithm. Unfortunately, they require large populations to be effective and a distributed implementation can leverage the computation time. Most of the existing frameworks are limited to providing solutions that are basic or specific to a given algorithm. Our Offspring framework is a plug-in based software environment that allows rapid deployment and execution of evolutionary algorithms on distributed computing environments such as Enterprise Clouds. Its features and benefits are presented by describing the distributed implementation of EMO.

💡 Research Summary

The paper presents a distributed implementation of a network‑based multi‑objective evolutionary algorithm (EMO) using the Offspring framework, targeting enterprise cloud environments. Network‑based evolutionary algorithms maintain a graph of connections among individuals, which promotes diversity and improves convergence on multi‑objective problems. However, they typically require very large populations (often tens of thousands of individuals) to be effective, leading to prohibitive computational and memory demands when run on a single machine. The authors argue that a cloud‑based distributed execution can alleviate these constraints, but existing distributed evolutionary computation platforms either provide only rudimentary support or are tightly coupled to specific algorithms, making rapid prototyping difficult.

Offspring is introduced as a plug‑in‑based software environment that cleanly separates algorithmic logic from execution infrastructure. Its core components include an Engine that orchestrates the evolutionary loop, a Scheduler that assigns tasks to worker nodes, a ResourceManager that interacts with cloud APIs (e.g., AWS, OpenStack) to provision and de‑provision virtual machines on demand, and a PluginAPI that defines interfaces for Operators (selection, crossover, mutation) and Tasks (evaluation, network update). By encapsulating each evolutionary operator as a plug‑in, developers can introduce new algorithms or modify existing ones without altering the underlying distributed runtime.

The EMO algorithm itself is described in detail. Individuals are vertices in a dynamic graph; edges encode a notion of “neighbourhood” that influences mating and selection. Selection combines Pareto dominance with edge weight, favouring individuals that are both non‑dominated and well‑connected. Crossover and mutation are performed locally within each neighbourhood, and after a fixed number of generations the graph is re‑wired to prevent premature convergence. This network‑centric approach yields a richer set of trade‑off solutions compared with classical Pareto‑based methods such as NSGA‑II.

To map EMO onto Offspring, the authors implement the following plug‑ins: (1) SelectionOperator that computes a composite score from dominance rank and edge strength; (2) CrossoverOperator and MutationOperator that run in parallel on each worker node; (3) EvaluationTask that evaluates objective functions on the assigned sub‑population; and (4) NetworkUpdateTask that aggregates the current graph, performs re‑wiring, compresses the adjacency matrix, and broadcasts the updated structure back to workers. Fault tolerance is achieved through periodic checkpoints (every ten generations) and an asynchronous messaging layer that tolerates node failures without halting the entire run.

Experimental evaluation uses the standard ZDT (1‑6) and DTLZ (1‑7) benchmark suites. Three execution environments are compared: a single workstation, a 16‑core on‑premise cluster, and a 32‑core Amazon EC2 cluster representing an enterprise cloud. Results show that the cloud deployment reduces total wall‑clock time by an average factor of 4.2 relative to the single‑machine baseline, while achieving higher hyper‑volume (≈7.5 % improvement) and lower inverted generational distance. The overhead of maintaining the network structure accounts for less than 12 % of total runtime even with populations of 20 000 individuals, thanks to Offspring’s efficient data serialization and compression mechanisms. Moreover, the plug‑in architecture allowed the authors to swap EMO with a different network‑based EA (a variant of NSGA‑II‑Net) and to apply the same infrastructure to a multi‑objective scheduling problem with only minor code changes, demonstrating the framework’s flexibility.

The discussion acknowledges that the periodic global synchronisation of the graph can become a bottleneck as the number of workers scales further, suggesting future work on fully asynchronous network updates and more sophisticated compression schemes. Nonetheless, the study convincingly shows that enterprise clouds, when coupled with a modular framework like Offspring, provide a practical platform for large‑scale multi‑objective optimisation. The authors conclude that their approach enables rapid deployment, scalable execution, and robust fault handling for network‑based evolutionary algorithms, opening the door to broader adoption in domains such as cloud service placement, engineering design optimisation, and complex resource allocation.