Defensive formations in professional Counter-Strike 2 (CS2) encode multi-agent spatiotemporal coordination, yet how these formations evolve has rarely been modeled. This paper proposes TAGNet, a Temporal Attention Graph Network that recognizes the defensive tactical patterns of the Counter-Terrorist (CT) side without tactical labels. TAGNet models the five CT players at five snapshots (10–20 s after freeze time) as fully connected graphs, encodes each snapshot with a 3-layer distance-aware Graph Attention Network (GAT) trained by self-supervised reconstruction, and aggregates the snapshot embeddings and per-snapshot 25-dimensional context vectors (economy, equipment, utility, score, and previous round) with a fixed, order-aware Gated Recurrent Unit (GRU) into 64-dimensional round embeddings. On 697 CT defensive rounds from 33 top-tier Dust2 matches (HLTV, 2024–2025), TAGNet achieves a test-set reconstruction loss of 0.0086, 8.5% below an earlier two-layer, three-snapshot configuration. Multi-metric K-Means clustering ($K{=}6$) discovers six interpretable defensive patterns, from anchor defense (C0) to aggressive forward pressure (C5). CT economy type is the contextual factor most strongly associated with tactical selection ($\chi ^{2}{=}96.25$, $p{< }0.001$, Cramér's $V{=}0.215$), and aggressive forward pressure is used significantly more often when trailing ($p{< }0.01$). The patterns also yield team-level tactical fingerprints for pre-match scouting. Ablations show that the graph encoder is essential for reconstruction and that the temporal sequence and context inputs make the patterns more interpretable; six robustness checks, including a rule-based baseline and Holm-corrected testing, support these choices. To our knowledge, this is the first unsupervised temporal graph learning framework for CT-side defensive tactical analysis in professional CS2 on Dust2.
This study formulates feasibility testing for partially revealed Minesweeper states as a quadratic unconstrained binary optimization (QUBO) problem. Binary decision variables are assigned only to hidden tiles, while revealed tiles and optionally fixed mines are treated as constants. For each revealed clue, a constraint is constructed matching the sum of adjacent hidden-mine variables to an effective clue value that accounts for any fixed neighboring mines, and we optionally include a global mine-count constraint over the unknown region. Squaring the constraint residuals yields a nonnegative objective whose zero-energy assignments correspond to mine placements consistent with the observed state. This study also derives closed-form expressions for the resulting QUBO coefficients, clarifying how the global mine-count term induces dense uniform couplings and how local clues add additional sparse structure. A 9-by-9 toy instance demonstrates coefficient construction. The formulation is compatible with classical optimization heuristics and quantum optimization pipelines that operate on QUBO/Ising models.
In multiplayer online games, intelligent matchmaking systems are the core scheduling mechanisms responsible for dynamically balancing player demands and system states. However, the traditional rule-based matchmaking system tries to balance matching quality and time consumption by manually adjusting multiple preset parameters, which often has poor real-time adaptability in response to the dynamic fluctuations of players. To make a smarter system, a “predict-then optimize” matchmaking (PTOM) framework is proposed, which integrates a Feature-interactive Behavior Sequential Transformer (FBST) prediction model and a Lagrangian dual-based dynamic programming algorithm. FBST uses a monotonicity constraint (ensuring that longer waiting times are associated with higher success rates) to predict the matching success rate based on real time player online data, while the optimization algorithm can use the prediction results to automatically obtain the best solution to multiple preset parameters through constructing and solving an optimization problem. Offline experiments demonstrate that FBST outperforms baselines in prediction accuracy and mono tonicity preservation, while the optimization algorithm achieves higher rewards and smaller duality gaps. A/B tests on online game (20 days, per-click randomization, reported per mode) show that the framework significantly improves the matching success rate (by 2.77% in normal and 0.93% in ranking matches) and reduces average matching time (by 20.40% in normal and 7.99% in ranking matches), validating its effectiveness in handling real time high-dimensional matchmaking challenges. We release the code at https://anonymous.4open.science/r/PTOM-E4DB/.
Cybersickness remains a major barrier to virtual reality adoption in serious games, yet audio-based mitigation is underexplored compared with visual interventions. This pa per presents an anticipatory audio architecture that predicts upcoming motion events and generates procedural sound cues before the visual motion occurs, implemented with MetaSounds in Unreal Engine 5. The design is motivated by predictive coding accounts of motion sickness, in which advance auditory information about impending motion may help the brain form more accurate sensory predictions. Technical evaluation on a single representative maritime scene shows motion-to-audio latency below 15 milliseconds with modest computational overhead. We report an exploratory cross-sectional study with 105 participants. Because all participants experienced the same anticipatory-audio configuration and no control condition was included, the study cannot establish the causal efficacy of anticipatory audio; it tests associations, not effects. Participants who perceived higher audio quality reported less motion sickness; the association was carried mainly by perceived audio–visual synchronization, survived demographic adjustment, and attenuated to marginal significance once session duration and cognitive load were controlled. Physiological measures corroborated the sickness self-report rather than the intervention. We conclude with design guidelines labelled as empirically supported or engineering-derived and outline the controlled comparison needed to establish efficacy.
Mass-casualty incident (MCI) training requires diverse scenarios while maintaining instructional consistency. However, most existing digital triage simulations remain script driven, relying on manually authored patient cases and predefined injury assets that limit scalability and replayability. This paper proposes a hybrid procedural serious game framework that combines LLM-based semantic patient synthesis with deterministic START validation to generate controllable training scenarios. Given predefined scenario specifications, structured patient profiles are procedurally generated and integrated into a two-layer architecture that separates semantic patient generation from geometric scene construction. To support safe-aware visual generation, validated injury descriptors are transformed into safety-constrained prompts for stylized low-poly visual generation while preserving triage-relevant cues. Experimental results demonstrate controllable variability across different seeds, along with repeatability of patient records and scene configurations under fixed conditions. Generation reliability did not improve consistently with model size. A four-week randomized study further shows that repeated gameplay improves triage decision speed relative to the control condition, while providing no significant additional gain in accuracy. These findings show that LLMs can function as controllable semantic generators within hybrid procedural frameworks, supporting replayable MCI training while retaining rule-based triage validation and safety aware visual design.
