Esport - Prediction using a neural network¶

This jupyter notebook can be found here https://drive.google.com/file/d/1pSjUtsf3tacWbv0bdXDcfGVdhrHg4pJx/view?usp=sharing

  • Demo file: https://drive.google.com/file/d/1AVZ3IdrN-TCkclVUcSEyjwrYMH58aqWe/view?usp=sharing

Project requirements:
numpy
pandas
scikit-learn
keras
matplotlib
seaborn
demoparser2
jupyterlab
tensorflow

Useful links:

  • HLTV results - https://www.hltv.org/results
  • CS2 events - https://cs2.poggu.me/dumped-data/game-events/
  • CS2 map sizes - https://github.com/pnxenopoulos/awpy/blob/main/awpy/data/map_data.py

CS2 Parser:

  • Parser - https://github.com/LaihoE/demoparser
  • Documentation - https://github.com/LaihoE/demoparser/tree/main/documentation/python
  • Example - https://github.com/LaihoE/demoparser/blob/main/examples/1vX/main.py
  • Examples - https://github.com/LaihoE/demoparser/tree/main/examples

Imports¶

In [1]:
from demoparser2 import DemoParser
import matplotlib.pyplot as plt
import matplotlib.image as mpimg
from matplotlib.collections import LineCollection
import numpy as np
import pandas as pd
from matplotlib import cm
from keras.models import Sequential
from keras.layers import Dense, Input
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler

Functions¶

Data overview¶

In [2]:
parser = DemoParser("demos/eternal-fire-vs-natus-vincere-m2-mirage.dem")

Events¶

In [3]:
df_deaths = parser.parse_event("player_death", player=["X", "Y"], other=["total_rounds_played"])
df_weapon_fires = parser.parse_event("weapon_fire", player=["X", "Y"], other=["total_rounds_played"])
df_player_spawns = parser.parse_event("player_spawn", other=["total_rounds_played"])
df_player_teams = parser.parse_event("player_team", other=["total_rounds_played"])
df_player_hurts = parser.parse_event("player_hurt", other=["total_rounds_played"])
df_round_starts = parser.parse_event("round_start")
df_round_ends = parser.parse_event("round_end")
df_last_round_half = parser.parse_event("round_announce_last_round_half")
df_item_purchases = parser.parse_event("item_purchase", other=["total_rounds_played"])
df_item_pickups = parser.parse_event("item_pickup", other=["total_rounds_played"])
df_bomb_plants = parser.parse_event("bomb_planted", other=["total_rounds_played"])
df_bomb_defuses = parser.parse_event("bomb_defused", other=["total_rounds_played"])
df_bomb_explosions = parser.parse_event("bomb_exploded", other=["total_rounds_played"])
df_bomb_drops = parser.parse_event("bomb_dropped", other=["total_rounds_played"])
df_bomb_pickups = parser.parse_event("bomb_pickup", other=["total_rounds_played"])
df_hegrenade_detonations = parser.parse_event("hegrenade_detonate", other=["total_rounds_played"])
df_flashbang_detonations = parser.parse_event("flashbang_detonate", other=["total_rounds_played"])
df_smokegrenade_detonations = parser.parse_event("smokegrenade_detonate", other=["total_rounds_played"])
df_smokegrenade_expirations = parser.parse_event("smokegrenade_expired", other=["total_rounds_played"])
df_inferno_startburns = parser.parse_event("inferno_startburn", other=["total_rounds_played"])
df_inferno_expirations = parser.parse_event("inferno_expire", other=["total_rounds_played"])
df_player_pings = parser.parse_event("player_ping", other=["total_rounds_played"])
df_player_sounds = parser.parse_event("player_sound", other=["total_rounds_played"])
df_ticks = parser.parse_ticks(["X", "Y"])
In [4]:
max_round = df_player_spawns["total_rounds_played"].max() + 1
max_round
Out[4]:
18

Team interpretation in numbers:

  • 3 = CT
  • 2 = T
In [5]:
df_player_teams
Out[5]:
disconnect isbot oldteam silent team tick total_rounds_played user_name user_steamid
0 False False 2 True 3 87273 12 jL 76561198176878303
1 False False 3 True 2 87273 12 Calyx 76561198052546265
2 False False 2 True 3 87273 12 w0nderful 76561199063068840
3 False False 3 True 2 87273 12 Wicadia 76561198812513923
4 False False 3 True 2 87273 12 MAJ3R 76561197967432889
5 False False 2 True 3 87273 12 Aleksib 76561198013243326
6 False False 3 True 2 87273 12 XANTARES 76561198044118796
7 False False 2 True 3 87273 12 b1t 76561198246607476
8 False False 2 True 3 87273 12 iM 76561198050250233
9 False False 3 True 2 87273 12 woxic 76561198083485506

Another approach

In [6]:
df_team_info2 = parser.parse_player_info();
df_team_info2
Out[6]:
steamid name team_number
0 76561198176878303 jL 3
1 76561198052546265 Calyx 2
2 76561199063068840 w0nderful 3
3 76561198812513923 Wicadia 2
4 76561197967432889 MAJ3R 2
5 76561198013243326 Aleksib 3
6 76561198044118796 XANTARES 2
7 76561198246607476 b1t 3
8 76561198050250233 iM 3
9 76561198083485506 woxic 2

Predicting of player deaths in a game round¶

In [7]:
df_weapon_fires
Out[7]:
silenced tick total_rounds_played user_X user_Y user_name user_steamid weapon
0 False 1258 0 1296.000000 32.000000 Aleksib 76561198013243326 weapon_smokegrenade
1 False 1514 0 -1372.368164 -977.939453 XANTARES 76561198044118796 weapon_knife_karambit
2 False 1661 0 -1921.021362 -325.519287 Calyx 76561198052546265 weapon_knife_karambit
3 False 1723 0 386.088837 51.468979 b1t 76561198246607476 weapon_flashbang
4 True 1766 0 -1174.323120 -727.842529 XANTARES 76561198044118796 weapon_usp_silencer
... ... ... ... ... ... ... ... ...
3066 False 137623 17 -2260.312012 628.897278 Aleksib 76561198013243326 weapon_deagle
3067 False 137650 17 -2255.038330 627.889587 Aleksib 76561198013243326 weapon_deagle
3068 True 137667 17 -1801.680664 -54.884399 iM 76561198050250233 weapon_m4a1_silencer
3069 True 137674 17 -1800.396851 -55.175659 iM 76561198050250233 weapon_m4a1_silencer
3070 False 137684 17 -1847.084473 -98.833923 w0nderful 76561199063068840 weapon_awp

3071 rows × 8 columns

Aggregating Data - Simplified example¶

Overview of weapon fires dataframe.

In [8]:
df_weapon_fires
Out[8]:
silenced tick total_rounds_played user_X user_Y user_name user_steamid weapon
0 False 1258 0 1296.000000 32.000000 Aleksib 76561198013243326 weapon_smokegrenade
1 False 1514 0 -1372.368164 -977.939453 XANTARES 76561198044118796 weapon_knife_karambit
2 False 1661 0 -1921.021362 -325.519287 Calyx 76561198052546265 weapon_knife_karambit
3 False 1723 0 386.088837 51.468979 b1t 76561198246607476 weapon_flashbang
4 True 1766 0 -1174.323120 -727.842529 XANTARES 76561198044118796 weapon_usp_silencer
... ... ... ... ... ... ... ... ...
3066 False 137623 17 -2260.312012 628.897278 Aleksib 76561198013243326 weapon_deagle
3067 False 137650 17 -2255.038330 627.889587 Aleksib 76561198013243326 weapon_deagle
3068 True 137667 17 -1801.680664 -54.884399 iM 76561198050250233 weapon_m4a1_silencer
3069 True 137674 17 -1800.396851 -55.175659 iM 76561198050250233 weapon_m4a1_silencer
3070 False 137684 17 -1847.084473 -98.833923 w0nderful 76561199063068840 weapon_awp

3071 rows × 8 columns

Overview of deaths dataframe.

In [9]:
df_deaths
Out[9]:
assistedflash assister_X assister_Y assister_name assister_steamid attacker_X attacker_Y attacker_name attacker_steamid attackerblind ... total_rounds_played user_X user_Y user_name user_steamid weapon weapon_fauxitemid weapon_itemid weapon_originalowner_xuid wipe
0 False NaN NaN None None 296.651398 -1586.485962 Aleksib 76561198013243326 False ... 0 150.391434 -1526.738159 Wicadia 76561198812513923 glock 17293822569165815812 20660576547 0
1 False NaN NaN None None -452.761658 -1758.071045 Aleksib 76561198013243326 False ... 0 -531.607178 -2378.742432 XANTARES 76561198044118796 glock 17293822569165815812 20660576547 0
2 False -531.607178 -2378.742432 XANTARES 76561198044118796 -1570.978027 -760.632751 MAJ3R 76561197967432889 False ... 0 -1272.865112 -1050.418091 iM 76561198050250233 usp_silencer 17293822569145761853 13626608964 0
3 False NaN NaN None None -812.001831 -2082.587646 w0nderful 76561199063068840 False ... 0 -465.970337 -2350.779785 Calyx 76561198052546265 glock 17293822569105195012 34398553971 0
4 False NaN NaN None None NaN NaN None None False ... 0 -1693.296631 -1375.607422 Aleksib 76561198013243326 planted_c4 0
... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
120 False NaN NaN None None -1698.991211 682.782166 jL 76561198176878303 False ... 17 -1064.924072 151.871338 XANTARES 76561198044118796 m4a1_silencer 17293822569131868220 34241486015 0
121 False -2065.616211 -559.346924 iM 76561198050250233 -1699.885620 681.705627 jL 76561198176878303 False ... 17 -1210.908081 285.837585 MAJ3R 76561197967432889 m4a1_silencer 17293822569131868220 34241486015 0
122 False -1064.924072 151.871338 XANTARES 76561198044118796 -1022.358459 120.646042 woxic 76561198083485506 False ... 17 -1699.831299 683.073303 jL 76561198176878303 ak47 17293822569170534407 25065970543 0
123 False -1699.792603 683.121765 jL 76561198176878303 -2495.580322 328.931458 Aleksib 76561198013243326 False ... 17 -1021.273499 123.897125 woxic 76561198083485506 deagle 17293822569105129473 29398974889 0
124 False NaN NaN None None -1847.084473 -98.833923 w0nderful 76561199063068840 False ... 17 -1736.028687 849.204041 Calyx 76561198052546265 awp 17293822569146089481 36570875316 0

125 rows × 33 columns

Merge dataframes for a round-level analysis.

In [10]:
df = pd.merge(df_weapon_fires, df_deaths, on=["total_rounds_played", "user_steamid"], how="left")
df["player_died"] = np.where(df["attacker_name"].notnull(), 1, 0)
df
Out[10]:
silenced tick_x total_rounds_played user_X_x user_Y_x user_name_x user_steamid weapon_x assistedflash assister_X ... tick_y user_X_y user_Y_y user_name_y weapon_y weapon_fauxitemid weapon_itemid weapon_originalowner_xuid wipe player_died
0 False 1258 0 1296.000000 32.000000 Aleksib 76561198013243326 weapon_smokegrenade False NaN ... 5866.0 -1693.296631 -1375.607422 Aleksib planted_c4 0.0 0
1 False 1514 0 -1372.368164 -977.939453 XANTARES 76561198044118796 weapon_knife_karambit False NaN ... 4059.0 -531.607178 -2378.742432 XANTARES glock 17293822569165815812 20660576547 0.0 1
2 False 1661 0 -1921.021362 -325.519287 Calyx 76561198052546265 weapon_knife_karambit False NaN ... 4449.0 -465.970337 -2350.779785 Calyx glock 17293822569105195012 34398553971 0.0 1
3 False 1723 0 386.088837 51.468979 b1t 76561198246607476 weapon_flashbang NaN NaN ... NaN NaN NaN NaN NaN NaN NaN NaN NaN 0
4 True 1766 0 -1174.323120 -727.842529 XANTARES 76561198044118796 weapon_usp_silencer False NaN ... 4059.0 -531.607178 -2378.742432 XANTARES glock 17293822569165815812 20660576547 0.0 1
... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
3127 False 137623 17 -2260.312012 628.897278 Aleksib 76561198013243326 weapon_deagle NaN NaN ... NaN NaN NaN NaN NaN NaN NaN NaN NaN 0
3128 False 137650 17 -2255.038330 627.889587 Aleksib 76561198013243326 weapon_deagle NaN NaN ... NaN NaN NaN NaN NaN NaN NaN NaN NaN 0
3129 True 137667 17 -1801.680664 -54.884399 iM 76561198050250233 weapon_m4a1_silencer NaN NaN ... NaN NaN NaN NaN NaN NaN NaN NaN NaN 0
3130 True 137674 17 -1800.396851 -55.175659 iM 76561198050250233 weapon_m4a1_silencer NaN NaN ... NaN NaN NaN NaN NaN NaN NaN NaN NaN 0
3131 False 137684 17 -1847.084473 -98.833923 w0nderful 76561199063068840 weapon_awp NaN NaN ... NaN NaN NaN NaN NaN NaN NaN NaN NaN 0

3132 rows × 40 columns

Neural network¶

nn.png

Data selection¶

  • We use data with _x suffix, because we would like to use name of the attacker, victim and weapon from the df_weapon_fires dataframe.
In [11]:
X = df[["user_X_x", "user_Y_x", "weapon_x", "total_rounds_played"]]
y = df["player_died"]
In [12]:
X
Out[12]:
user_X_x user_Y_x weapon_x total_rounds_played
0 1296.000000 32.000000 weapon_smokegrenade 0
1 -1372.368164 -977.939453 weapon_knife_karambit 0
2 -1921.021362 -325.519287 weapon_knife_karambit 0
3 386.088837 51.468979 weapon_flashbang 0
4 -1174.323120 -727.842529 weapon_usp_silencer 0
... ... ... ... ...
3127 -2260.312012 628.897278 weapon_deagle 17
3128 -2255.038330 627.889587 weapon_deagle 17
3129 -1801.680664 -54.884399 weapon_m4a1_silencer 17
3130 -1800.396851 -55.175659 weapon_m4a1_silencer 17
3131 -1847.084473 -98.833923 weapon_awp 17

3132 rows × 4 columns

In [13]:
y
Out[13]:
0       0
1       1
2       1
3       0
4       1
       ..
3127    0
3128    0
3129    0
3130    0
3131    0
Name: player_died, Length: 3132, dtype: int32

Data processing¶

One-hot encode the 'weapon_x' column (since it's categorical).

In [14]:
X = pd.get_dummies(X, columns=["weapon_x"])
X
Out[14]:
user_X_x user_Y_x total_rounds_played weapon_x_weapon_ak47 weapon_x_weapon_awp weapon_x_weapon_deagle weapon_x_weapon_elite weapon_x_weapon_flashbang weapon_x_weapon_galilar weapon_x_weapon_glock ... weapon_x_weapon_knife_widowmaker weapon_x_weapon_m4a1_silencer weapon_x_weapon_mac10 weapon_x_weapon_molotov weapon_x_weapon_mp9 weapon_x_weapon_p250 weapon_x_weapon_smokegrenade weapon_x_weapon_ssg08 weapon_x_weapon_tec9 weapon_x_weapon_usp_silencer
0 1296.000000 32.000000 0 False False False False False False False ... False False False False False False True False False False
1 -1372.368164 -977.939453 0 False False False False False False False ... False False False False False False False False False False
2 -1921.021362 -325.519287 0 False False False False False False False ... False False False False False False False False False False
3 386.088837 51.468979 0 False False False False True False False ... False False False False False False False False False False
4 -1174.323120 -727.842529 0 False False False False False False False ... False False False False False False False False False True
... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
3127 -2260.312012 628.897278 17 False False True False False False False ... False False False False False False False False False False
3128 -2255.038330 627.889587 17 False False True False False False False ... False False False False False False False False False False
3129 -1801.680664 -54.884399 17 False False False False False False False ... False True False False False False False False False False
3130 -1800.396851 -55.175659 17 False False False False False False False ... False True False False False False False False False False
3131 -1847.084473 -98.833923 17 False True False False False False False ... False False False False False False False False False False

3132 rows × 27 columns

Scale data to improve the performance and convergence.

In [15]:
scaler = StandardScaler()
X_scaled = scaler.fit_transform(X)
X_scaled
Out[15]:
array([[ 2.58971678,  0.94432862, -1.73770765, ..., -0.03575993,
        -0.09834216, -0.27419013],
       [-0.83079472, -0.09453845, -1.73770765, ..., -0.03575993,
        -0.09834216, -0.27419013],
       [-1.53409894,  0.57656893, -1.73770765, ..., -0.03575993,
        -0.09834216, -0.27419013],
       ...,
       [-1.38111921,  0.85495559,  1.7833376 , ..., -0.03575993,
        -0.09834216, -0.27419013],
       [-1.37947353,  0.85465599,  1.7833376 , ..., -0.03575993,
        -0.09834216, -0.27419013],
       [-1.43932117,  0.80974723,  1.7833376 , ..., -0.03575993,
        -0.09834216, -0.27419013]])

Data split for training and testing phases with training size of 80 %.

In [16]:
X_train, X_test, y_train, y_test = train_test_split(X_scaled, y, test_size=0.2, random_state=42)
In [17]:
X_train.shape
Out[17]:
(2505, 27)

Model with three dense layers and one input layer with size of X_train.

  • Input layer is used to define the input shape.
  • Last layer uses sigmoid activation because of binary classification.
In [18]:
model = Sequential()
model.add(Input(shape=(X_train.shape[1],)))
model.add(Dense(64, activation="relu"))
model.add(Dense(32, activation="relu"))
model.add(Dense(1, activation="sigmoid"))

Compile the model.

In [19]:
model.compile(loss="binary_crossentropy", optimizer="adam", metrics=["accuracy"])
model.summary()

Model: "sequential"

┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━┓
┃ Layer (type)                         ┃ Output Shape                ┃         Param # ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━┩
│ dense (Dense)                        │ (None, 64)                  │           1,792 │
├──────────────────────────────────────┼─────────────────────────────┼─────────────────┤
│ dense_1 (Dense)                      │ (None, 32)                  │           2,080 │
├──────────────────────────────────────┼─────────────────────────────┼─────────────────┤
│ dense_2 (Dense)                      │ (None, 1)                   │              33 │
└──────────────────────────────────────┴─────────────────────────────┴─────────────────┘

 Total params: 3,905 (15.25 KB)

 Trainable params: 3,905 (15.25 KB)

 Non-trainable params: 0 (0.00 B)

Model training phase.

In [20]:
model.fit(X_train, y_train, epochs=10, batch_size=32, validation_split=0.1)

Epoch 1/10
71/71 ━━━━━━━━━━━━━━━━━━━━ 1s 3ms/step - accuracy: 0.6496 - loss: 0.6540 - val_accuracy: 0.7092 - val_loss: 0.5485
Epoch 2/10
71/71 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.7318 - loss: 0.5323 - val_accuracy: 0.7450 - val_loss: 0.5223
Epoch 3/10
71/71 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.7652 - loss: 0.4948 - val_accuracy: 0.7649 - val_loss: 0.5166
Epoch 4/10
71/71 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.7619 - loss: 0.4936 - val_accuracy: 0.7729 - val_loss: 0.5036
Epoch 5/10
71/71 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.7986 - loss: 0.4503 - val_accuracy: 0.7849 - val_loss: 0.4979
Epoch 6/10
71/71 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.7780 - loss: 0.4663 - val_accuracy: 0.7809 - val_loss: 0.4909
Epoch 7/10
71/71 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.7807 - loss: 0.4568 - val_accuracy: 0.7610 - val_loss: 0.4928
Epoch 8/10
71/71 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.7864 - loss: 0.4481 - val_accuracy: 0.7610 - val_loss: 0.4833
Epoch 9/10
71/71 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.7803 - loss: 0.4657 - val_accuracy: 0.7928 - val_loss: 0.4674
Epoch 10/10
71/71 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.7958 - loss: 0.4403 - val_accuracy: 0.7769 - val_loss: 0.4839
Out[20]:
<keras.src.callbacks.history.History at 0x1b2451c5880>

Testing phase by evaluating the model with testing data.

In [21]:
loss, accuracy = model.evaluate(X_test, y_test)
print(f'Model Accuracy: {accuracy}')

20/20 ━━━━━━━━━━━━━━━━━━━━ 0s 790us/step - accuracy: 0.7759 - loss: 0.4630
Model Accuracy: 0.7735247015953064

Task 2¶

Create two different neural networks to predict some interesting situations that may affect the game.

Use the previous example for implementation.

Here are some examples:

  • Predicting bomb plant in a game round - Predict if a bomb will be planted in a given round based on weapon fires, player positions, and past round data.
  • Predicting the team to win a game round - Predict which team will win the round based on player actions and round data.
  • Predicting player damage in a game round - Predict the total damage a player will deal in a round based on their actions and positioning.
  • Predicting flashbang detonations effectiveness - Predict the effectiveness of a flashbang based on the number of players it flashes in a round.