CS2 Role Consistency & Round Outcomes

Measuring behavioural player roles in Counter-Strike 2 with Self-Organizing Maps —
and testing whether staying "in role" actually helps you win the round.

Bachelor's Thesis Project · Sachin Kumar (S20240010206) · IIIT Sri City

---

An end-to-end research pipeline that

1. derives behavioural player roles from professional CS2 demo files using per-side Self-Organizing Maps (following Drachen, Canossa & Yannakakis, IEEE CIG 2009),
2. quantifies each player's role consistency across a match, and
3. tests whether prior consistency predicts round outcomes under a leakage-free, prospective design with match-clustered inference.

flowchart LR
    A["raw .dem files"] --> B["parse<br/>rounds · 2 Hz ticks · events"]
    B --> C["19 behavioural<br/>features / player-round"]
    C --> D["per-side SOM<br/>role discovery"]
    D --> E["role consistency<br/>leakage-free rolling"]
    E --> F["round outcome<br/>match-clustered logit"]
    style A fill:#6E4AA5,color:#fff
    style D fill:#E8710A,color:#fff
    style F fill:#2E7D32,color:#fff

Loading

Headline result

Consistency and winning travel together within a half (OLS β = 0.54, p = 0.002) — but a team's prior average consistency does not predict the next round (logit β = 0.32, 95% CI [−0.88, 1.52]).

The null survives eco-exclusion, taxonomy granularity (k ± 1), per-map splits, and a 100-replicate permutation test. The raw association appears to run substantially winning → consistency, not the reverse: winning preserves the economy, which preserves the plan, which preserves roles.

_{Left: round-outcome logit — the positive control (equipment) is strongly significant while prior consistency is a tight null. Right: win rate is flat across prior-consistency quartiles (48.4 → 46.7%).}

The role taxonomy

Roles are discovered, not assumed — a SOM is trained per side and the codebook is clustered into archetypes, which are then named by hand from their behavioural signatures. K_BY_SIDE = {T: 4, CT: 5}, locked for this corpus.

Side	Roles
T	Aggressor · AWPer · Support Rifler · Lurker
CT	Site Anchor · AWPer · Rotating Rifler · Eco/Save · Aggressor

_{Top-10 discriminative features per role (z-scored vs the side mean). Radars carry the final role names, so they match the report exactly.}

More figures — SOM U-matrices, consistency distribution, smoke-test traces

_{U-matrices show the trained 22×22 self-organizing maps; the smoke-test traces are the round-1 positional sanity check against the official HLTV scoreline.}

Corpus

Maps	76 professional maps
Events	10 (2024–2025) — majors, IEM, ESL Pro League
Teams	13
Rounds	1,652
Player-rounds	16,520 (8,260 per side)
Determinism	fixed seed (42) + fixed config reproduce every thesis number

Raw .dem files are HLTV/Valve content and are not redistributed — manifest.csv is the dataset specification for re-downloading the exact corpus, and scripts/extract_dems.ps1 closes the acquisition loop.

Repository layout

cs2-btp/
├── manifest.csv              source of truth: one row per demo + status
├── code/cs2btp/              the Python package (9 modules)
├── notebooks/                7 Colab driver notebooks (run in order)
├── models/                   role profiles & names (persisted taxonomy)
├── analysis/                 consistency tables, regression datasets, QC,
│                             stability, permutation & robustness artifacts
├── figures/                  thesis-ready PNGs (200 dpi)
├── docs/                     the reference documentation set
├── scripts/                  extract_dems.ps1 (dataset acquisition)
├── raw_dems/     (gitignored) input .dem files (event__stage__match naming)
├── parsed/       (gitignored) per-demo parquet (rounds, ticks, events)
└── features/     (gitignored) player-round feature tables (+ role labels)

Quickstart (Google Colab + Drive)

1. Upload code/cs2btp/ to My Drive/cs2-btp/code/cs2btp/ and notebooks/ to My Drive/cs2-btp/notebooks/; put demos in raw_dems/.
2. Open each notebook in Colab and run in order (table below). Every stage is resumable — re-running skips finished work.
3. After replacing any .py file in Drive: delete code/cs2btp/__pycache__ and Runtime → Disconnect and delete runtime before re-running (Colab caches imported modules aggressively).

#	Notebook	Stage	Typical runtime
00	setup_and_smoke_test	env check; parse ONE demo; validate vs HLTV	~5 min
01	parse_all_demos	parse everything in the manifest (resumable)	1–3 min/demo
02	build_features	19 behavioural features per player-round (cached)	~1 min/demo
03	discover_roles	per-side SOMs; choose k; name roles; stability	45–70 min
04	consistency_metrics	modal share / entropy / switch; rolling variant	~2 min
05	outcome_analysis	round-level logit; descriptives; half-level OLS	~2 min
06	robustness	no-eco; k ± 1; per-map; permutation; weakest-link	60–90 min

Notebook 03 contains the two human-in-the-loop cells (K_BY_SIDE, FINAL_NAMES). Both are locked with the final decisions for this corpus and safe to Run-all.

Documentation (docs/)

File	Contents
ARCHITECTURE_AND_DATA.md	pipeline architecture, data flow, every file schema
MODULE_REFERENCE.md	API reference for all 9 modules + full config reference
NOTEBOOK_GUIDE.md	per-notebook inputs/outputs, editable cells, checks
DECISIONS_AND_RESULTS.md	the complete decision log + results summary
REPRODUCTION_AND_TROUBLESHOOTING.md	exact reproduction steps, environment, known issues

Start with PROJECT_CONTEXT.md for the self-contained ground-truth overview.

Requirements

Google Colab (CPU is sufficient) with Drive mounted. Installed by the bootstrap cell: demoparser2 (pinned 0.41.3), minisom. Pre-installed on Colab: pandas, numpy, scipy, scikit-learn, statsmodels, joblib, matplotlib, pyarrow. See requirements.txt.

Methodological lineage

Role discovery replicates Drachen, Canossa & Yannakakis, "Player Modeling using Self-Organization in Tomb Raider: Underworld" (IEEE CIG 2009): behavioural telemetry → emergent SOM → cluster the map → manual inspection and naming → stability analysis. This is distinct from Drachen, Sifa, Bauckhage & Thurau, "Guns, Swords and Data" (IEEE CIG 2012, large-scale behavioural clustering) — cite both, correctly attributed.

Citation

If you use this pipeline or its findings, please cite it via the repository's CITATION.cff (GitHub → Cite this repository):

Kumar, S. (2026). Player Role Consistency and Round Outcomes in Counter-Strike 2. Bachelor's Thesis, Indian Institute of Information Technology, Sri City.

License

Released under the MIT License for the pipeline code and notebooks. Derived statistical artifacts (features, models, analysis tables) are provided for research reproducibility only; source .dem files are HLTV/Valve content and are not distributed.