starcraft winner prediction
Post on 13-Apr-2017
32 Views
Preview:
TRANSCRIPT
StarCraft Winner PredictionPresented by: Yaser Norouzzadeh
Authors: Y. Norouzzadeh, S. Bakkes, P. Spronck
Tilburg Center for Cognition and CommunicationTilburg University
Tilburg, the Netherlands
2
Table of Contents• Introduction• Data• Features• Method• Results• Conclusion
Y.NorouzzadehRavari@uvt.nl
3
Table of Contents• Introduction• Data• Features• Method• Results• Conclusion
Y.NorouzzadehRavari@uvt.nl
4
StarCraft Environment
Y.NorouzzadehRavari@uvt.nl
Player B basePlayer A base
5
StarCraft Races
Y.NorouzzadehRavari@uvt.nl
6
StarCraft Match Types
Y.NorouzzadehRavari@uvt.nl
TvZ
PvZPvT
7
Winner Prediction• Complexity:
• many action choices• managing units concurrently• different strategies• various match types (PvT, PvZ, TvZ, PvP, ZvZ, and TvT)
• Winner:• The first player who destroy all enemy units
• Application:• Evaluation function for AI bots
Y.NorouzzadehRavari@uvt.nl
8
Table of Contents• Introduction• Data• Features• Method• Results• Conclusion
Y.NorouzzadehRavari@uvt.nl
9
Dataset Information• Expert replays collected by Synnaeve and Bessiere 2012• Database provided by Robertson and Watson 2014 • Filter condition:
If (game-length < 10 min) or (game-length > 50 min) or (game-winner == Null) Remove replay
Match Type PvT PvZ TvZ PvP ZvZ TvT
Number of replays 2017 840 812 392 199 395
Number of replays(After filtering) 1490 579 612 263 115 298
Y.NorouzzadehRavari@uvt.nl
10
Winning Rate
PvT PvZ TvZ48%
49%
50%
51%
52%
53%
54%
55%
56%
57%
Wining rate percent in non-symmetric match types
Y.NorouzzadehRavari@uvt.nl
11
Map Types• 60% of maps have a size of 128 × 128 tiles• Different maps are used
Y.NorouzzadehRavari@uvt.nl
12
Table of Contents• Introduction• Data• Features• Method• Results• Conclusion
Y.NorouzzadehRavari@uvt.nl
13
Feature sets• Time-Dependent Features (TDF)• Time-Independent Features (TIF)
Sample size per match typeMatch types PvT PvZ TvZ PvP ZvZ TvT
Feature samples 24k 9k 9k 3k 1k 4k
Y.NorouzzadehRavari@uvt.nl
14
Extracting Time-Dependent Features
10 seconds Features during 10s
180 seconds TDF(mean,var,dif)
Y.NorouzzadehRavari@uvt.nl
15
Time-Dependent Features• Proposed features:• Number of frequent commands (move, build, tech, hold, siege, and burrow)• Number of micro/macro commands• Number of control/strategy/tactic commands• Number of unique regions that include a building• Difference of building values of all regions
• Well-known features:• Unspent resources: available resources on average at any given time• Income: total resources that are collected over T• APM
Y.NorouzzadehRavari@uvt.nl
16
Extracting Time-Independent Features
Map TIF
Y.NorouzzadehRavari@uvt.nl
17
Time-Independent Features (TIF)• Number of regions• Ratio of buildable tiles• Ratio of walkable tiles• Average of choke distances• Height level ratio (low, low doodads, high, high doodads, very high,
very high doodads)• Map dimension in tiles
Y.NorouzzadehRavari@uvt.nl
18
Table of Contents• Introduction• Data• Features• Method• Results• Conclusion
Y.NorouzzadehRavari@uvt.nl
19
Problem Formulation• Binary classification: win(1)/lose(0)• Classification methods• Gradient Boosting Regression Trees (GBRT)• Random Forest (RF)
• Approaches:• Individual model for each match type (6 models)• Mixed models
• For symmetric match types (PvP, ZvZ, and TvT)• For non-symmetric match types (PvT, PvZ, and TvZ)
• General model for all match types
Y.NorouzzadehRavari@uvt.nl
20
Table of Contents• Introduction• Data• Features• Method• Results• Conclusion
Y.NorouzzadehRavari@uvt.nl
21
A=APM and economy featuresB=time-dependent featuresC=time-independent features
Winner Prediction by Individual models
Model Features PvT PvZ TvZ PvP ZvZ TvT
baseline 0.55 0.51 0.56 0.50 0.50 0.50RF A,B,C 0.591 0.611 0.502 0.502 0.515 0.491
GBRT A,B,C 0.595 0.623 0.502 0.502 0.507 0.483
RF A,B 0.644 0.634 0.624 0.643 0.587 0.639
GBRT A,B 0.637 0.634 0.624 0.639 0.581 0.635
Y.NorouzzadehRavari@uvt.nl
22
Winner Prediction by Mixed Models and General Model
Model Features Non-symmetric Symmetric General
RF A,B,C 0.575 0.497 0.591
GBRT A,B,C 0.577 0.499 0.593
RF A,B 0.639 0.637 0.639
GBRT A,B 0.635 0.634 0.635
Y.NorouzzadehRavari@uvt.nl
23
Top-10 TDF In Individual Models
• Income and unspent resources always amongst top three (Except ZvZ)• In ZvZ, micro commands have strongest predictive value• Control commands (move, gather, build, …) and region values are the next strongest predictive value
Y.NorouzzadehRavari@uvt.nl
24
Top-10 TDF In Mixed/General Model
Y.NorouzzadehRavari@uvt.nl
25
Table of Contents• Introduction• Data• Features• Method• Results• Conclusion
Y.NorouzzadehRavari@uvt.nl
26
Conclusion• Winner prediction is possible for all match types.• Mixed models also manage to predict the match winner as individual
models.• In all models, top-10 features are more or less the same.• Economic features (income and unspent) are strongest features across
match types
Y.NorouzzadehRavari@uvt.nl
27
Thank you for your attention.
Y.NorouzzadehRavari@uvt.nl
top related