Year: 2026
Type: preprint
Abstract: Draft-phase recommendation in multiplayer online battle arena (MOBA) games must adapt to balance patches that repeatedly shift the game meta. In shared-weight continual learning, adapting to a new patch can overwrite parameters that remain important for earlier patches, creating a fundamental stability-plasticity trade-off. We propose a patch-routed mixture-of-experts architecture that isolates adaptation across patches while retaining a shared representation. Each patch is assigned a dedicated expert selected through its patch identity, which is naturally available at deployment in versioned games. This patch-aware routing prevents cross-patch interference while allowing new patches to adapt without modifying earlier patch-specific knowledge. Across our experiments, the proposed architecture matches sequential fine-tuning's plateau at the first patch transition while reaching it from roughly half as many post-patch samples, and achieves zero forgetting on earlier patches at 63% of the per-patch storage required to maintain an independent model for every patch. The zero-forgetting property is verified by weight-identity checks against saved models at every patch and seed. The speed advantage holds over the first two patch transitions and is lost as the stream lengthens, scoping the architecture to streams short relative to the frozen trunk's useful lifetime.
Source: Research Square
Institution: Semnan University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204108480