Community

Clash Royale and Machine Learning

Inspired by GTA and Machine Learning, I want ask some hints to develop a AI that can actually play Clash Royale. Now I trying to develop a huge dataset with every game details.

player,card,startTime,endTime,px,py
1,26000016,114,134,9,31
1,28000001,201,221,0,31
0,26000012,273,293,14,12
0,26000007,302,322,13,10
1,26000013,346,366,11,27
1,26000001,511,531,17,29
0,26000015,662,682,14,13
1,26000000,673,693,14,17
0,26000014,771,791,14,10
0,26000003,961,981,14,12
1,28000000,984,1004,11,24
1,28000001,1115,1135,17,0
0,26000001,1133,1153,13,13
0,26000012,1387,1407,10,18
0,26000007,1561,1581,10,18

This is an example of a .csv file that can describe an entire match. In a more human format it means something like that:

gadm_banana  VS  Mendax3.0
________________________________
Mendax3.0 plays prince px:9 py:31 at 114 ticks
Mendax3.0 plays arrows px:0 py:31 at 201 ticks
gadm_banana plays skeleton-army px:14 py:12 at 273 ticks
gadm_banana plays witch px:13 py:10 at 302 ticks
Mendax3.0 plays bomber px:11 py:27 at 346 ticks
Mendax3.0 plays archers px:17 py:29 at 511 ticks
gadm_banana plays baby-dragon px:14 py:13 at 662 ticks
Mendax3.0 plays knight px:14 py:17 at 673 ticks
gadm_banana plays musketeer px:14 py:10 at 771 ticks
gadm_banana plays giant px:14 py:12 at 961 ticks
Mendax3.0 plays fireball px:11 py:24 at 984 ticks
Mendax3.0 plays arrows px:17 py:0 at 1115 ticks
gadm_banana plays archers px:13 py:13 at 1133 ticks
gadm_banana plays skeleton-army px:10 py:18 at 1387 ticks
gadm_banana plays witch px:10 py:18 at 1561 ticks

(px, py are the position of the tiles of the battlefield and ticks it's the time: 20 ticks == 1 second)

It's think that a good architecture to apply to this kind of problems are Recurrent Neural Network(RNN). I've already work with LSTM (a type of RNN) but I can't figure out how to apply RNN to this problem. Should I pass all the data together or split between players ?  What's the best way to approach this project?