controlstack questions

Prompt

# V13 Population-to-Target Player Transfer Tournament You are designing a **bounded experiment**, not a new architecture, for a CS2 professional-player mimicry project. The core unresolved question is how to use a broad high-Elo human demo corpus when the actual objective is imitation of **one particular professional player**. Do not assume more population data is always better. Do not assume target-only training is always better. Design the smallest experiment that can distinguish useful transferable structure from regression toward population-average behavior. ## Current project state A large normalized high-Elo/FACEIT corpus already exists. A reconstructed UserCmd action authority has now been qualified for supported demos. It can expose human command supervision including: - W/A/S/D-equivalent held states; - movement fields; - view angles; - mouse deltas/counts; - subtick input-history; - modifiers/actions; - command/player identity. A candidate first canonical authority contains 362 approved demos and ~489M reconstructed commands, pending materialization. The actual project objective is not generic strong CS2 behavior. It is **mimicry of a target professional player**. ## Current v13 architecture The initial canonical path is direct semantic action imitation from causal human-admissible observations. At a high level: ```text admissible observation/context → optional predictive/tactical representation → human semantic action distribution → coherent selected action → deterministic human-equivalent realizer → Source 2 ``` The canonical planar locomotion action boundary uses four held W/S/A/D-equivalent bits, with modifiers/actions and continuous view represented separately as appropriate. The project will begin with simple direct-action baselines before promoting more complicated prospective/tactical representations. ## Known risk A broad population corpus may improve: - motor/action priors; - representation learning; - rare-state coverage; - general tactical structure; - stability. But it may also erase exactly what we want: - target-specific timing; - route preference; - movement style; - spacing; - aim dynamics; - aggressiveness; - transition hazards; - unusual but stable target-player habits. A model can improve average held-out likelihood while becoming a worse mimic of the target. ## Candidate strategies worth comparing You may consider, but are not limited to: 1. target-only training; 2. population pretraining then target fine-tuning; 3. shared representation with a target-specific policy head; 4. player-conditioned population model; 5. population model followed by target-specific adapters/LoRA-like small modules; 6. population prior with explicit target residual; 7. representation pretraining on population while training the action head only on target data. These are examples, not mandatory arms. ## Important experimental requirement The experiment must distinguish: ```text genuine useful transfer ``` from: ```text simply adding more data / regularization ``` and from: ```text regression toward high-Elo population mean ``` ## Data-split constraints Avoid leakage across: - matches; - temporally adjacent rounds where relevant; - duplicated/reprocessed demos; - target-player appearances inside population pretraining if an arm claims target exclusion. Where possible, preserve harder temporal/tournament/map splits to test whether target-specific style generalizes. The target player's available data may be much smaller than the population corpus, so sample-efficiency matters. ## Evaluation philosophy Do not rely only on action cross-entropy. Target mimicry should examine distributions such as: - W/A/S/D occupancy; - held-state dwell times; - transition and reversal hazards; - counter-strafe/overlap behavior; - movement speed/state distributions after Source 2 realization; - view angular velocity/acceleration/jerk; - aim corrections/overshoot where measurable; - fire relative to movement/aim state; - spatial/nav occupancy and transitions; - nearest-teammate/centroid positioning; - opening/rotation/engagement tendencies; - conditional behavior by map/side/weapon/game state. The exact metric subset should stay bounded. ## Your task Design a **minimal population→target transfer tournament** that canonical ML could realistically run after the direct-action baseline exists. The tournament should answer: 1. Does population data improve target-player imitation? 2. Which parts of the model should transfer? 3. Does population training wash out target-specific behavior? 4. Is any gain merely regularization/sample-size rather than transferable representation? 5. How much target data is needed to recover target identity/style? ## Required output ### 1. Minimal experiment arms Choose the smallest set of arms needed to answer the questions above. For each arm specify: - what is trained on population data; - what is trained on target data; - what is frozen/reinitialized; - whether target-player data is excluded from pretraining. Avoid redundant variants. ### 2. Data-budget curve Propose target-data fractions or sample budgets sufficient to measure sample efficiency without an expensive sweep. ### 3. Evaluation Give: - primary target-player action metric(s); - a small process-level style metric set; - population-regression diagnostics; - calibration/probabilistic metrics if needed. Explain how to tell: - better generic modeling; - better target mimicry; - worse target identity despite improved NLL. ### 4. Transfer attribution Design at least one control that distinguishes: - representation transfer; - motor/action prior transfer; - ordinary regularization from extra data. ### 5. Decision rules State evidence that would justify: - target-only canonical training; - population pretraining + target adaptation; - shared/player-conditioned representation; - deferring the question because evidence is ambiguous. Do not choose based on elegance or parameter count alone. ### 6. Failure interpretations If population pretraining hurts target mimicry, distinguish plausible causes such as: - distribution mismatch; - over-regularization; - insufficient target adaptation; - target-specific tactical context; - model capacity; - metric failure. Keep the output experiment-oriented. Do not redesign the broader v13 stack.