Logo image
A fine-grained actor-based autoencoder model with fault tolerance
Dissertation   Open access

A fine-grained actor-based autoencoder model with fault tolerance

Sithembiso Dyubele
Doctor of Philosophy (PHD), University of Johannesburg
2026
Handle:
https://hdl.handle.net/10210/520039

Abstract

The exponential growth in data volumes and computational complexity of modern artificial intelligence systems has exposed fundamental limitations in traditional distributed computing approaches. Existing frameworks suffer from scalability bottlenecks, inadequate fault tolerance mechanisms, communication safety vulnerabilities, and architectural designs that resist effective distribution across heterogeneous computing environments. This study presents a comprehensive framework that addresses these challenges by applying principled actor-based computing to distributed systems and machine learning architectures. With its inherent properties of message passing communication, fault isolation, and location transparency, the actor model provides a unifying theoretical foundation for building resilient, scalable, and type-safe distributed systems. Main Contributions: This thesis makes five interconnected research contributions demonstrated across chapters 8 to 12, culminating in chapter 7’s unified actor-based machine learning platform. First, we established theoretical foundations proving that actor-based distribution achieves optimal fault isolation while maintaining convergence guarantees. Second, we developed type-safe communication protocols that eliminate runtime errors through compile-time verification. Third, we designed component-level fault-tolerance mechanisms achieving 99.4% recovery success with a 2.3-second mean recovery time. Fourth, we demonstrated architecture-agnostic distribution across variational autoencoders, LSTM networks, and hybrid optimisation algorithms. Fifth, we validated scalability from single-node to 16-node deployments with quantified performance characteristics. Key Advantages: The actor-based approach provides three fundamental advantages over traditional distributed computing paradigms. Fault tolerance through isolation: Each actor operates independently, enabling the system to recover from individual component failures without cascading effects—achieving 99.4% automatic recovery compared to manual intervention required by conventional approaches. Type-safe communication: Compile-time verification eliminates the entire class of runtime communication errors that plague distributed systems, achieving zero message delivery errors across 10,000+ test executions. Architecture-agnostic scalability: The message-passing abstraction enables consistent scaling characteristics (86.4% efficiency at 8 actors, 71% at 16 actors) across diverse neural architectures without architecture-specific optimisation. 2 Quantified Benefits: The unified platform delivered substantial performance improvements across all evaluation metrics. Computational efficiency improved with 98% execution time reduction (24.5s to 0.50s) while maintaining model quality. Feature selection achieved 80% dimensionality reduction with 98.6% performance preservation, enabling deployment on resource-constrained devices. The fault-tolerant architecture maintained 100% message delivery success with automatic recovery from node failures, eliminating the 4-8 hour manual intervention typical in production systems. Hybrid optimisation achieved 88.1% fitness scores, outperforming genetic algorithms (77.3%) and particle swarm optimisation (71.8%) on complex search spaces. Most significantly, the platform demonstrated consistent 86.4% scaling efficiency with eight actors and 71% at 16 actors, substantially exceeding the 50-60% efficiency typical of data-parallel frameworks at similar scales. Practical Impact: These results enable three immediate practical applications. First, organisations can deploy resilient distributed ML systems that automatically recover from hardware failures without human intervention. Second, the type-safe communication protocols eliminate a significant source of production bugs, accelerating development cycles. Third, the architecture-agnostic design allows practitioners to apply actor-based distribution to new neural architectures without reimplementation, thereby lowering the barrier to adopting distributed training. The experimental evaluation demonstrated consistent improvements across all implementations in performance metrics (64.5% training time reduction), fault tolerance (99.4% recovery success, 2.3-4.2s recovery time), communication efficiency (43% overhead reduction, 7.2% average overhead), and scalability (zero synchronisation barriers, complete parallelisation) compared to state-of-the-art baseline methods, including PyTorch Distributed Data Parallel, Horovod, and traditional distributed frameworks.
pdf
Dyubele, S.4.11 MBDownloadView
Open Access

Metrics

1 Record Views

Details

Logo image