PUBG Fortune Teller

吳易倫   李政諺

Motivation

  • Compare with friends
  • Discover talent
  • Improve gaming ability

Data collection

Sample nicknames of players

PUBG official API

Third party API

User status Database

Query resource id of player

Get stats of each player per month

Data collection

Solo(單排) Duo(雙排) Squad(四排)
TPP(第三人稱) Solo-TPP Duo-TPP Squad-TPP
FPP(第一人稱) Solo-FPP Duo-FPP Squad-FPP

Data collection

Sample Query id Get stats
# of players 70000 43000 22000
May Mar, Apr ... May, Apr...Jan
# of players  19000 ~19000 ... 12659

# of players who have played more than 5 matchs

# of players at each step

Data collection

Attribute Name Description Mean
​damage_dealt_avg 每場平均傷害量 122.16
kills_sum 擊殺數量 71.05
matches_cnt 總場數 81.16
topten_cnt 前10名場數 8.13
assists_sum 協助擊殺 5.11
headshot_kills_sum 爆頭總數 15.03
time_survived_avg 每場平均存活時間 715.34

Correlation among attributes

Problem Formulation

  • Input: a ​sequence of stats in the previous months

  • Output: ​damage_dealt_avg in next few months

Models

  • Lasso
  • Linear Regression
  • DNN
  • LSTM+DNN
  • AutoEncoder+DNN
  • Lasso
  • Linear Regression
  • DNN
  • LSTM+DNN
  • AutoEncoder+DNN

Lasso

Least Absolute Shrinkage and Selection Operator

{1\over 2}\Vert y-XW\Vert_2^2+\alpha\Vert W\Vert_1^1+\gamma\sum_{t=1}^{m-1}\Vert w_{t+1}-w_t\Vert
12∥y−XW∥22+α∥W∥11+γ∑t=1m−1∥wt+1−wt∥{1\over 2}\Vert y-XW\Vert_2^2+\alpha\Vert W\Vert_1^1+\gamma\sum_{t=1}^{m-1}\Vert w_{t+1}-w_t\Vert
  • The optimization objective for Lasso is:
X:=\text{input stats matrix},X\in R^{N\times n}
X:=input stats matrix,X∈RN×nX:=\text{input stats matrix},X\in R^{N\times n}
n:=\text{the size of concatenated stats vector},n\in N
n:=the size of concatenated stats vector,n∈Nn:=\text{the size of concatenated stats vector},n\in N
m:=\text{the number of forecast value}, m\in N
m:=the number of forecast value,m∈Nm:=\text{the number of forecast value}, m\in N
y:=\text{target label matrix},y\in R^{N\times m}
y:=target label matrix,y∈RN×my:=\text{target label matrix},y\in R^{N\times m}
W:=\text{weighting matrix},W\in R^{n\times m}
W:=weighting matrix,W∈Rn×mW:=\text{weighting matrix},W\in R^{n\times m}
W=[w_1,w_2,...,w_m],w_i\in R^{n\times 1}
W=[w1,w2,...,wm],wi∈Rn×1W=[w_1,w_2,...,w_m],w_i\in R^{n\times 1}
N:=\text{the size of training data}
N:=the size of training dataN:=\text{the size of training data}

Lasso

The characteristics of objective function:

  • L1 penalty 
    • Reduce training time
    • Robust
    • feature selection
  • L2 penalty
    • ​Stable
{1\over 2}\Vert y-XW\Vert_2^2+\alpha\Vert W\Vert_1^1+\gamma\sum_{i=1}^{n-1}\Vert w_{i+1}-w_i\Vert_1^1
12∥y−XW∥22+α∥W∥11+γ∑i=1n−1∥wi+1−wi∥11{1\over 2}\Vert y-XW\Vert_2^2+\alpha\Vert W\Vert_1^1+\gamma\sum_{i=1}^{n-1}\Vert w_{i+1}-w_i\Vert_1^1

Lasso

The third term is useless

{1\over 2}\Vert y-XW\Vert_2^2+\alpha\Vert W\Vert_1^1+\gamma\sum_{i=1}^{n-1}\Vert w_{i+1}-w_i\Vert_1^1
12∥y−XW∥22+α∥W∥11+γ∑i=1n−1∥wi+1−wi∥11{1\over 2}\Vert y-XW\Vert_2^2+\alpha\Vert W\Vert_1^1+\gamma\sum_{i=1}^{n-1}\Vert w_{i+1}-w_i\Vert_1^1

NN models

DNN

LSTM+DNN

LSTM-Autoencoder+DNN

Results:single target

Model MSE Baseline
Lasso 4067 5987
Linear Regression 4088 5987
DNN 4402 5987
LSTM+DNN 4417 5987
LSTM-Autoencoder+DNN 4460 5987

Train input: 1, 2, 3     Train output: 4

Test input: 2, 3, 4       Test output: 5

Results:single target

Train input Train output Test input Test output MSE Baseline
1 2 2 3 4044 5751
2 3 3 4 3855 5470
1,2 3 2,3 4 3783 5471
3 4 4 5 4468 5987
2,3 4 3,4 5 4177 5987
1,2,3 4 2,3,4 5 4067 5987

Lasso

Results

damage_dealt_avg

kills_avg

time_survived_avg

Results

damage_dealt_avg

kills_avg

time_survived_avg

Results

damage_dealt_avg

kills_avg

time_survived_avg

Results:Multi-target

Train input Train output Test input Test output MSE Baseline
2 3,4 3 4,5 4148 5825
1,2 3,4 2,3 4,5 4122 5825

Results:Multi-target

Results:Multi-target

Demo

Future Work

  • Collect more data
  • Recommend teammates

Charts

Charts

Made with Slides.com