Starcraft II Build Order Classification

Contents

• Proposal

  • AlphaStar plays really well (executes actions close to flawlessly), however, the type of strategies it employs are limited and does not adapt to opponent moves especially in the late game.
  • Informed MCTS is an interesting strategy to enable exploitation of possible enemy action distribution.
  • However, we first require some prediction of enemy actions. Here, Bayesian classification and multivariate regression can help.
  • First cluster buildorder sequences to inform priors.
  • Next feed information from clustering into regression for next possible step taken.

• Data Visualization

  • spawningtool extracted data: Discrete action list of units, structures, and upgrades produced. Unfortunately does not make available game-level resource information such as minerals available or supply utilization.
  • sc2reader parsing helper functions adapted from IBM/starcraft2-replay-analysis under Apache 2.0 License. sc2reader helps us with the extraction of resource information. In particular, we are interested in the distribution of resources across the 3 classes {economy, army, technology} as the game progresses.

• Feature Extraction

  • Time Series extracted information is a huge dictionary for each player's statistics in a match and follows the following structured format. The 'data' field is a pandas dataframe with each column as a distinct feature in this particular instance of a time series (1 game for 1 player).

  '1597403940_2305463': {
       'race': 'P',
       'matchup': 'PvP',
       'data':      mineral_collection_rate  mineral_per_worker_rate  mineral_queued_army  \
       0                          0                 0.000000                    0   
       1                        293                24.416667                    0   
       2                        671                51.615385                    0   
       3                        671                51.615385                    0   
       4                        755                53.928571                    0   
       ..                       ...                      ...                  ...   
       121                     1903                32.254237                    0   
       122                     1875                31.779661                    0   
       123                     1903                33.982143                    0   
       124                     1903                35.240741                    0   
       125                     1903                35.240741                    0 
  ...
  

  • Action Sequence extracted information (pending compression & cleanup) largely follows the same format.

  '1597403940_2305463': {
       'race': 'P',
       'matchup': 'PvP',
       'data': {
            [...] # Raw spawningtool 'buildOrder' field dictionary for now => can be compressed
       }
  }
  

  • As the data is quite large, we use python's (>=3.0) in-built compression tool bzip2 to reduce our footprint (to around 10MB from a 350MB replay pack - of course, we also threw away quite some information along the way). Each set of extracted data can be easily extracted and loaded into a python dictionary using the included zipUtil functions:

  from zipUtil import zip_write, zip_read
  # Write
  dic = {} # some dictionary
  zip_write('filename', dic)
  # Read
  read_dic = zip_read('filename')

• State-Action Generation

  • The nearest slice of timeseries information is found for each action taken and put into a (state,action) feature-label pair. In addition to the timeseries information (on economic resources), accumulated actions taken thus far is also extracted as a set of discrete features.

• Naive Bayes + Baselines

  • Baseline: Random Forest Classifier & AdaBoost
  • Naive Bayes: Condition on continuous, discrete, mixture + KDE likelihood estimation
  • Gaussian Process Classification: Binary One-vs-Rest -> softmax
    • Hard to select inducing set, and did not perform above heuristic strategy of picking most frequent actions
  • Results are put into a json and plotted here

Results

Top-1 and Top-3 label classification accuracy is tested and computed on a 70/10/20 train/validation/test split between the selected models. Unfortunately, we only average a measly ~10% improvement over the naive heuristic option of picking k-most frequent actions in order.

1. Features are highly correlated, Naive Bayes also doesn't do so well beyond top-1
2. Dataset may be underdetermined, in the sense that the same (for discrete features) or close feature vectors are labeled differently.

[P]	Baseline (RF)	Heuristic	Continuous (Multinomial)	Discrete (Complement)	Mixture	KDE (exp)
Top-1	0.412	0.399	0.399	0.421	0.420	0.444
Top-3	0.626	0.586	0.604	0.587	0.621	0.678

[T]	Baseline (RF)	Heuristic	Continuous (Multinomial)	Discrete (Complement)	Mixture	KDE (exp)
Top-1	0.426	0.352	0.410	0.424	0.427	0.446
Top-3	0.668	0.644	0.649	0.584	0.623	0.707

[Z]	Baseline (RF)	Heuristic	Continuous (Multinomial)	Discrete (Complement)	Mixture	KDE (exp)
Top-1	0.396	0.322	0.346	0.411	0.409	0.442
Top-3	0.638	0.596	0.608	0.648	0.669	0.716

Purpose

Automate clustering of replays into different types of Build Orders in an unsupervised and possibly online fashion.

Resources

Dataset: 1979 Replays (Patch >= 5.0.2 BU)

• Pro-players, non-AI games

Curated Icon set: Structure,Unit,Upgrades Icon Set + File mapping

• Just unzip into same directory as the notebooks and it should work
• Python data stuctures used for mapping, will update to more portable .json format at a later time
• All copyrights belong to Blizzard Entertainment Inc., fair usage for purposes of scientific publication/educational use in this project.