Source-linked AI summary
HoopMind: A Real-Time Neural Game-Tree System for Opponent-Aware Possession Planning
Yibo Gong, Cong Guo, Jiacheng Ding
TL;DR
School coaches lack the analytics tools available to professional teams, and the paper asks how far public data can close that gap. HoopMind fuses public basketball data and uses learned shot values in a pruned possession planner; pruning reduces search computation while usually retaining the exhaustive search’s top action.
Problem
School coaches prepare for opponents with film and intuition, while professional-team analytics tools remain out of reach; the paper asks how public data can help close this gap.
Source: §I · p. 1Method
HoopMind fuses public sources into per-shot data, estimates opponent profiles and calibrated shot probabilities, then searches a half-court decision tree with depth-limited expectimax and branch-and-bound pruning.
Source: §I · p. 1, B. Pruning and complexity · p. 3Results
43% fewer nodes were evaluated, with the same top action as exhaustive search in 97% of states at k=6 and ε=0.02.
Source: B. Pruning and complexity · p. 3Takeaways & Limitations
The authors expect the public-data pipeline to carry over to other invasion sports with open event data, naming soccer and ice hockey as next ports.
Source: C. Beyond basketball · p. 5, §VII · p. 5The pruning bound is a heuristic, not provably admissible, so pruning is approximate rather than guaranteed exact.
Source: B. Pruning and complexity · p. 3
Abstract
from arXiv · showhide
School coaches prepare for opponents with game film and intuition. The analytics tools of professional teams stay out of reach. We ask how far public data can close this gap. Professional basketball is our case study, chosen for its data rather than the league. We fuse five public sources into one per-shot dataset of 4.23M shots over 21 seasons. The sources are shot locations, two play-by-play feeds, official matchup tracking, and player biometrics. Alignment across them is 99.5% to 100%. We also report two data pitfalls that are easy to miss. We then model a half-court possession as a sequential game. Shot values come from ShotNet, an embedding multilayer perceptron (MLP). On a held-out season it beats a zone-rate baseline and a logistic baseline, and its probabilities are well calibrated. A depth-limited expectimax search then solves the offensive decision tree, with branch-and-bound pruning to keep it real time. All training runs offline, so the online system stays light. A scouting planner and a playable simulator both run in a single browser page.
I. INTRODUCTION
HoopMind addresses coaches’ lack of accessible, data-grounded opponent preparation by combining public professional-basketball data with a real-time planning tool. Its contributions are a fused per-shot dataset, opponent profiles and ShotNet, and a browser-based planner and simulator.
Motivation
- The motivating gap is that school coaches rely on film, whiteboards, and intuition while professional-team analytics remain inaccessible to them.The proposed tool is intended to show where, against whom, and when an opponent is weak, and let players experience the answer by playing.Source: §I · p. 1
Data contribution
- The fused dataset aligns sources at 99.5% to 100%, with 99.9% agreement on matched outcomes, and includes pre-shot margin, assists, clutch status, blockers, and shooter biometrics.The authors also report two pitfalls that can corrupt a naive merge.Source: §I · p. 1
Models
- ShotNet is an embedding MLP trained on 3.8M shots; on a held-out season it beats zone-rate and logistic baselines under a strict time split.The work also uses empirical-Bayes shrinkage for rate estimates and reports matchup-data bias.Source: §I · p. 1
Planning and system
- A depth-limited expectimax search with branch-and-bound pruning plans half-court possessions, while two browser tools provide scouting and a playable simulator.The simulator uses one in-browser ShotNet forward pass per shot, and coach mode displays the live solution.Source: §I · p. 1
II. RELATED WORK
Prior work studies spatial shot patterns, defensive behavior, shot selection, and interactive basketball systems, but key planning approaches often rely on private optical tracking. HoopMind focuses on what can be built and evaluated with public data, including interactive deployment.
Spatial and tracking models
- Closest defensive-skill, possession-value, and ghost-defender approaches require private optical tracking, motivating a test of what similar goals can achieve with public data alone.The cited works are Franks et al., Cervone et al., and Le et al..Source: §II · p. 1
Shot outcome prediction
- ShotNet sits within the reported public-feature AUC band of about 0.60 to 0.68, so the paper emphasizes fused features, time-based evaluation, and interactive deployment rather than higher AUC.Adding closest-defender distance raises accuracy, but that feature is private.Source: §II · pp. 1–2
Decision value and interactive systems
- HoopMind places expected-points shot selection inside a searchable tree tied to a named opponent, while prior play-sketching work shares the interactive goal but uses private tracking.Commercial games use finite-state machines, influence maps, and steering; HoopMind replaces editor-set player ratings with estimates checked on a held-out season.Source: §II · p. 2
III. POSSESSION AS A PRUNED SEQUENTIAL GAME
The paper models a half-court possession as a zero-sum stochastic game, then fixes the defense to an empirical policy so the offense can be planned against it. A depth-limited expectimax search compares available actions, with explicit chance nodes for turnovers and learned estimates at terminal states.
A. Formulation
- A half-court possession is modeled as a two-team zero-sum stochastic game whose state includes player positions, ball location, attacker-defender distances, clocks, and score margin.The offense selects from actions such as shooting, driving, passing, screening, and cutting; the defense selects a coverage.Source: A. Formulation · p. 2
- Fixing the defense to an empirical policy estimated from data leaves a Markov decision process for the offense, making planning well defined.The policy tables encode what teams concede by zone, defender attackability, and defensive intensity by quarter.Source: A. Formulation · p. 2
- Contest and block uncertainty are priced in terminal payoffs, while drive and pass turnovers are represented as explicit chance nodes.A drive succeeds with probability 1 − λdr, set to 0.06, and a pass with probability 1 − λps, set to 0.035; otherwise the possession ends with zero points.Source: A. Formulation · p. 2, Figure 2 · p. 2
- The planner limits search depth, keeps shooting available at every state, and compares the value of shooting now with values from expanded actions.At cutoff, every leaf is a real terminal; the caption identifies red terminal nodes as priced by learned estimators and the dashed mark as a branch-and-bound cut.Source: A. Formulation · p. 2, Figure 2 · p. 2
B. Pruning and complexity
The depth-2 expectimax search uses optimistic action bounds and branch-and-bound pruning to reduce computation while preserving the exhaustive search’s top action in most tested states.
B. Pruning and complexity
- At depth d=2, the exhaustive tree evaluates 57.0 leaf shots on average; deeper search cost more browser time without changing the chosen action.Source: B. Pruning and complexity · p. 2
- 43% fewer nodes and 97% top-action agreement at the shipped setting (k=6, ε=0.02) over 300 random states.The comparison is against exhaustive depth-2 expectimax.Source: Table I · p. 3, B. Pruning and complexity · p. 3
- The search orders candidate drives and passes by optimistic bounds, cutting a branch when its bound falls at least ε below the best value found.The bound is heuristic, not provably admissible, so pruning is approximate and its cost is measured rather than assumed exact.Source: B. Pruning and complexity · p. 3
A. Sources and alignment
The authors join five public sources to form a per-shot dataset spanning 21 seasons, with high shot-to-play-by-play alignment and very high agreement on matched outcomes.
A. Sources and alignment
- Across 21 seasons, shot-to-play-by-play match rates are 99.5% to 100%, and the two feeds agree on 99.9% of matched outcomes.The join key is game, player, quarter, minute-left, and second-left.Source: A. Sources and alignment · p. 3
- Shot charts cover 2003-04 through 2023-24; both play-by-play feeds and biometrics start in 1996-97, while matchup tracking starts in 2017-18.Source: A. Sources and alignment · p. 3
- The court zones comprise 14 range-and-side combinations, and backcourt heaves—0.2% of attempts—are excluded from analyses.Source: A. Sources and alignment · p. 3
B. Two pitfalls worth recording
The section documents two data pitfalls that distort shot-profile analyses and reports findings made possible by the fused data.
B. Two pitfalls worth recording
- Using 0-0 event scores as missing values inflates the apparent clutch-shot share from about 4% to about 7.5%.The clue is a jump at the 2017/2018 file boundary.Source: B. Two pitfalls worth recording · p. 3
- Using post-shot scores makes teams ahead by 16 or more appear to shoot 55.4% versus 40.3% when behind by 16 or more; correcting for the shot flattens the curve to 46.9%–48.6%.Every margin in the paper is therefore a pre-shot margin.Source: B. Two pitfalls worth recording · p. 3
C. What fusion unlocks
- In 2023-24, clutch FG% falls from 47.7% to 43.9%, with 3PT FG% falling from 36.9% to 31.7%.Clutch means the last five minutes with |margin| ≤5.Source: C. What fusion unlocks · p. 3
- Corner threes are assisted on 96% to 97% of makes versus 46% of mid-range makes; 9.0% of two-point attempts are blocked versus 0.9% of threes.Across 21 seasons, mid-range attempts fall from 35.6% to 11.2%, while three-point attempts rise from 18.7% to 39.5%.Source: C. What fusion unlocks · p. 3
V. LEARNED OPPONENT PROFILES
The learned profiles estimate defensive performance by team, zone, defender, and quarter, while the paper cautions that defender rankings reflect assignments rather than pure ability.
A. Team, defender, and time profiles
- For each team and each of 14 zones, the model estimates allowed FG% and block rate with empirical-Bayes shrinkage toward league rates using a 100-attempt prior.Source: A. Team, defender, and time profiles · p. 3
- Defender rankings measure points allowed per 100 partial possessions when targeted, shrunk over at least 400 possessions; assignment bias means this is not pure ability.Coaches may hide weak defenders on weak scorers, and conditioning on assignment offensive strength could sharpen the ranking.Source: A. Team, defender, and time profiles · pp. 3–4
- Quarter-level FG% allowed gives each team an intensity curve; Boston defends hardest in the first quarter, 3.1 percentage points below its own average.Source: A. Team, defender, and time profiles · p. 3
- Fig. 3 reports alignment, a pre-shot versus post-shot margin comparison, and held-out-season reliability; Fig. 4 contrasts first- and last-season shot-density landscapes.The Fig. 4 caption describes the mid-range plateau draining into a ridge along the arc.Source: Figure 3 · p. 4, Figure 4 · p. 4
B. ShotNet
ShotNet is an embedding MLP trained on earlier seasons and evaluated on held-out 2023–24 shots. Its probabilities are well calibrated, and Table II reports its shot-outcome prediction results.
B. ShotNet
- ShotNet combines four learned embeddings with nine numeric features in a 96-64-1 MLP with about 46k parameters.Training uses 3.79M shots through 2021–22, validation uses 217k shots from 2022–23, and the 2023–24 test has 218k shots.Source: B. ShotNet · p. 4
- 0.016 expected calibration error across 10 bins; ShotNet’s mean prediction is 47.4% against a 47.5% base rate.The held-out season contains 218k shots and was never touched during development.Source: B. ShotNet · p. 4, Table II · p. 4
- The trained network exports to 362 KB of JSON and runs in about 104 multiply-adds per browser forward pass.Training takes minutes on one workstation GPU, using AdamW, batch size 16384, and four epochs.Source: B. ShotNet · p. 4
A. Scouting planner
The scouting planner presents opponent-specific attack options, while the playable simulator updates a pruned game-tree solution as a possession develops. Its defender paths are behavior AI because movement tracking is not published.
A. Scouting planner
- The planner colors zones by attack value against league average, ranks defenders by attackability, and shows the selected player’s history and per-quarter intensity.For a named attacker, prediction combines his zone profile with the opponent’s in log-odds space.Source: A. Scouting planner · pp. 4–5
B. Playable simulator with a live-solved game tree
- In the Boston example, a contested pull-up three is worth 0.79 expected points, while driving middle and kicking to the weak-side shooter is worth 1.24.Coach mode lists the top three branches and draws the best branch on court as the play develops.Source: B. Playable simulator with a… · p. 5, Figure 5 · p. 5
- The simulator reprices teammate shots when a helper leaves an assignment and runs the pruned search every 0.25 seconds.Each shot uses one ShotNet forward pass plus a game layer for openness and contest distance; the interface shows those coefficients.Source: B. Playable simulator with a… · p. 5
- Defender paths are behavior AI driven by estimates, because movement tracking is not published.The browser tools are self-contained HTML files, and the public repository can rebuild the dataset, profile tables, and ShotNet from raw public files.Source: B. Playable simulator with a… · p. 5
C. Beyond basketball
The paper presents a public-data pipeline for opponent-aware possession planning and expects it to extend beyond basketball where public event data are available. Its conclusion reports high source alignment and computational savings from pruned search.
C. Beyond basketball
- The proposed transfer recipe fuses public event streams, estimates opponent defense, learns calibrated terminal-action values, and solves the offensive decision tree with pruned search.The authors identify soccer open data and public ice-hockey shot logs as planned ports.Source: C. Beyond basketball · p. 5
VII. CONCLUSION
- The pruned search cuts computation substantially and still selects the same action as exhaustive search most of the time.The conclusion also reports high alignment across the fused sources.Source: §VII · p. 5