💻 computer science

A Patch-Routed Mixture-of-Experts for Continual MOBA Draft Recommendation

This paper proposes a patch-routed mixture-of-Experts architecture for continual MOBA draft recommendation that leverages versioned patch identities to isolate adaptation, achieving zero forgetting and significantly faster convergence with reduced storage compared to independent models, though its speed advantage diminishes as the stream of patches lengthens.

Original authors: Mohammadreza Mohammadnejad, Morteza Dorrigiv, Farzin Yaghmaee

Published 2026-08-25

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Mohammadreza Mohammadnejad, Morteza Dorrigiv, Farzin Yaghmaee

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of competitive video games, the most critical decisions often happen before the first shot is fired. In games like Dota 2, two teams take turns banning and selecting characters, known as heroes, to build their lineup. This drafting phase is a high-stakes puzzle where every choice limits future options and sets the stage for victory. Because the game is constantly being updated by its developers to keep things fresh, the rules change frequently. These updates, called patches, tweak the strength of characters, the cost of items, and how abilities interact. A strategy that works perfectly today might be useless tomorrow. This creates a unique challenge for artificial intelligence: how can a computer learn to make good recommendations when the very definition of "good" is shifting beneath its feet?

This is the problem of continual learning. Imagine a student who must master a new subject every month, but the textbooks for the previous months are still relevant and must not be forgotten. If the student tries to learn the new material by simply overwriting their old notes, they lose the knowledge needed for the past exams. This tension between learning new things and remembering old ones is a fundamental hurdle in artificial intelligence. Researchers have long sought ways to help machines adapt to new data without erasing what they have already learned, a struggle known as the stability-plasticity trade-off. In the context of video games, where patches arrive on a fixed schedule and their identity is known in advance, this problem offers a rare opportunity to test a specific solution.

A team of researchers from Semnan University tackled this issue by building a new kind of recommendation system for Dota 2 drafting. Instead of trying to force a single, flexible brain to remember every version of the game at once, they designed a system that keeps a shared core of knowledge while assigning a dedicated specialist to each game update. Their approach, called PatchExpertFFN, treats the game's version number not as a mystery to be guessed, but as a clear key that unlocks the right set of tools. When the game updates, the system knows exactly which version is active and routes the decision-making to a specific expert module trained for that era, leaving the older experts untouched.

The researchers tested this idea using a massive dataset of professional matches spanning five different game versions. They compared their new system against two other approaches: a standard method that tries to learn everything in one continuous stream, and a "model zoo" where a completely separate, full-sized AI is saved for every single patch. The standard method struggled significantly; as it learned the new patches, it forgot how to play the old ones, a phenomenon known as catastrophic forgetting. The model zoo, on the other hand, never forgot anything because it kept a perfect copy of every past version, but it required storing a massive amount of data, essentially keeping a full library of brains for every update.

The new system struck a remarkable balance. By using a shared backbone that was trained only on the first patch and then frozen, and then attaching a smaller, specialized expert for each subsequent patch, the researchers achieved a result that was nearly as good as the model zoo but much more efficient. Their system retained the ability to make accurate recommendations for earlier patches with zero loss of memory, matching the performance of the separate models. At the same time, it required only about 63 percent of the storage space needed by the model zoo. This efficiency came from the fact that the system did not need to store the entire brain for every version, only the specific parts that changed.

The study also revealed how quickly the system could adapt. When a new patch arrived, the new expert learned the necessary adjustments much faster than the standard method, reaching its peak performance in roughly half the number of training examples. However, this speed advantage had a limit. As the stream of patches grew longer, the fixed shared core became less adequate for the newest versions, and the system's performance began to lag slightly behind the fully separate models. The researchers found that the system works best when the stream of updates is relatively short compared to the useful life of that initial shared core. If the game changes too many times, the frozen foundation eventually becomes too distant from the current reality, requiring a periodic refresh of the core itself.

What makes this discovery particularly significant is the mechanism behind it. The system does not need to guess which version of the game is being played; the game client simply tells it. This allows the system to hard-wire the connection between the game version and the correct expert, ensuring that learning for the new version never accidentally overwrites the knowledge of the old ones. The researchers verified this by checking the internal weights of the system, confirming that the memory of past patches remained exactly as it was, untouched by the training of new ones. This structural guarantee of zero forgetting is something that other methods, which rely on complex mathematical balancing acts, can only approximate.

The findings suggest a practical path forward for AI systems that must operate in environments with clear, labeled changes. Whether it is software updates, seasonal catalog changes, or shifting market conditions, if the identity of the change is known, a system can be built to isolate adaptation without sacrificing memory. The researchers noted that while their method is highly efficient, it is not a permanent fix for an endless stream of changes. The system is designed for streams that are short relative to the lifespan of the initial shared knowledge. For longer-term applications, the strategy would involve a scheduled maintenance policy where the shared core is retrained to catch up with the current state of the world, resetting the chain of specialists.

In the end, this work demonstrates that the solution to the problem of forgetting does not always require more complex algorithms or larger memories. Sometimes, the answer lies in how the system is organized. By respecting the boundaries of change and assigning specific roles to specific eras, the researchers created a system that learns quickly, remembers perfectly, and does so with a fraction of the storage cost of previous solutions. It is a reminder that in the face of constant change, knowing exactly what has changed can be the most powerful tool of all.

Technical Summary: A Patch-Routed Mixture-of-Experts for Continual MOBA Draft Recommendation

1. Problem Formulation

The paper addresses the challenge of Continual Learning (CL) in the context of Multiplayer Online Battle Arena (MOBA) game draft recommendations. Specifically, it targets the "stability-plasticity trade-off" inherent in shared-weight models when adapting to balance patches.

In games like Dota 2, developers release periodic patches that alter hero statistics, ability values, and item costs. These patches induce a domain-incremental shift: the task (recommending the next hero pick) remains constant, but the input distribution changes abruptly at known, labeled timestamps.

The Core Conflict: Standard continual learning methods (e.g., Elastic Weight Consolidation) or sequential fine-tuning on a shared model often suffer from catastrophic forgetting, where adapting to a new patch overwrites parameters critical for previous patches.
The Constraint: Retraining a model from scratch for every patch is computationally expensive and discards useful knowledge. Conversely, maintaining a separate "model zoo" (one full model per patch) guarantees zero forgetting but incurs prohibitive storage costs.
The Opportunity: Unlike typical concept drift where the boundary is inferred, MOBA patches provide an explicit, known identity (game version metadata) at both training and inference time. The paper asks whether this known identity can be used to route parameters directly, isolating adaptation without inferring task boundaries.

2. Methodology: PatchExpertFFN

The authors propose PatchExpertFFN, a Patch-Routed Mixture-of-Experts (MoE) architecture designed to isolate adaptation while retaining a shared representation.

Architecture Design:

Shared Trunk: The backbone consists of a Transformer-based model (specifically a BERT4Rec variant) where the token/position embeddings, self-attention sublayers, and pre-attention normalization layers form a frozen trunk. This trunk is trained only on the initial patch (Patch 0) and never updated thereafter.
Per-Patch Experts: The feed-forward (FFN) sublayers of every backbone layer are replaced by a library of per-patch experts.
Hard Routing: Routing is hard-coded based on the patch identity ( $t$ ). A sequence drafted under patch $t$ is processed exclusively by the expert $E_t$ and the corresponding output head $H_t$ . No learned gating network is required; the patch ID acts as the switch.
Initialization Strategy: For $t \ge 1$ , the new expert $E_t$ and head $H_t$ are copy-initialized from the immediately preceding patch's expert ( $E_{t-1}$ ), rather than from the frozen trunk. This provides a "warm start" that encodes prior knowledge without requiring retraining of the trunk.
Training Schedule:
- Patch 0: The trunk, $E_0$ , and $H_0$ are trained jointly.
- Patch $t > 0$ : The trunk and all previous experts/heads ( $j < t$ ) are frozen. Only the current pair $(E_t, H_t)$ is trained on the new patch's data.

Theoretical Properties:

Zero Backward Transfer: Because no later patch's training updates an earlier patch's expert, backward transfer is zero by construction.
Storage Efficiency: The architecture stores one frozen trunk plus one expert library/head per patch. The authors calculate that the per-patch storage increment is approximately 63% of a full independent model (specifically $\rho \approx 0.628$ ), as the trunk is shared.

3. Key Contributions

PatchExpertFFN Architecture: A novel MoE design that replaces FFN sublayers with per-patch experts, hard-routed by explicit patch identity. It eliminates the need for learned gates or task inference steps.
Verified Zero-Forgetting: The authors verify via weight-identity checks that the implementation achieves exact zero backward transfer against saved models, distinguishing the result from approximate retention methods.
Storage vs. Retention Trade-off: The paper quantifies the cost of zero forgetting. PatchExpertFFN achieves 98.7% of the mean plateau performance of a full "model zoo" (which stores a full model per patch) while requiring only 63% of the per-patch storage.
Expert-Drift Instrumentation: The authors introduce a metric to measure the displacement of trained experts from their copy sources, quantifying how much adaptation occurs beyond the warm start.

4. Experimental Results

Experiments were conducted on the DOTA-Draft dataset, covering five consecutive Dota 2 patches (7.35–7.39) with professional match data.

Plasticity and Speed:
- At the first patch transition (Exp1), PatchExpertFFN matches the plateau performance of sequential fine-tuning (Base) but reaches it from roughly half as many post-patch samples.
- At subsequent transitions (Exp2–Exp4), the routed arm maintains 97–99% of the sequential fine-tuning plateau but shows a widening performance deficit.
Retention:
- Routed Measurement: When evaluated using the correct patch-specific expert, backward transfer is exactly zero.
- Unrouted Ablation: When the routing is removed and all past patches are scored using the final patch's expert, performance degrades significantly (similar to the Base model), proving that the retention comes from routing, not just freezing the trunk.
Storage:
- The model zoo (full models per patch) requires linearly growing storage. PatchExpertFFN grows linearly as well but with a slope of 0.628 relative to the full model size.
Expert Drift:
- Drift analysis shows that adaptation is distributed across all layers rather than concentrated in specific heads or layers. The drift grows non-monotonically but increases overall, indicating that later patches must absorb both their own shifts and the growing distance from the fixed trunk.

5. Significance and Limitations (as claimed by the paper)

Significance:
The paper demonstrates that explicit environmental metadata (patch version) can be leveraged to achieve structural parameter isolation in continual learning. By hard-routing to dedicated experts, the system guarantees zero forgetting without the storage overhead of maintaining full independent models. The approach is particularly effective for short-to-medium length streams where the shared trunk remains adequate.

Limitations and Scope:

Plasticity Tax: The performance gap between PatchExpertFFN and sequential fine-tuning widens over time. This is attributed to the frozen trunk (trained on Patch 0) becoming increasingly misaligned with the distribution of later patches. The architecture is scoped to streams short relative to the trunk's useful lifetime.
Unseen Patches: The current design cannot handle patches for which no expert has been trained yet (no defined routing target).
Generalizability: The results are specific to the Dota 2 dataset, the BERT4Rec backbone, and five patches. The "plasticity tax" observed is specific to this stream.
Future Work: The authors suggest that a periodic trunk refresh (retraining the shared trunk and restarting the expert chain) could mitigate the plasticity tax, though this was not evaluated. They also note that learned or soft routing could handle unseen patches.

In summary, the paper presents a pragmatic solution for MOBA draft recommenders that trades a small, time-dependent performance deficit for a significant reduction in storage costs and a strict guarantee against catastrophic forgetting, leveraging the unique property of known, labeled game versions.