- General Info
- Technologies Used
- Features
- Setup
- Object Detection
- Control of bot
- Machine learning
- Programming AI
- Results Summary
- Project Status
- Room for Improvement
- Contact
- License
The idea for this project came to me after playing the online video game: "Heroes of the Storm™". In the game, players form into five-player teams and fight against another team. Team consists of characters of different roles. One of the tasks of a player is to control the situation on the map's lane, by killing minions and collecting experience globes. I wanted to automize this process and started a project, to create robot player (bot) that will do it for me, as best as possible and without external help.
As the project envolved, I divided it into following sections:
- Object detection
- Control of bot
- Machine learning - bot as an agent
- Programming AI
- Object detection: opencv, win32com
- Control of bot: pyautogui
- Machine learning: keras, tensorflow, gym, json
- Detecting in-game objects for further evaluation
- Controlling movements and actions of the bot to gather game's experience points
- Executing escape maneuvers when the bot is at risk af dying
- Create the account on: battle.net
- Download and install the game: "Heroes of the Storm™"
- Start game's launcher, start game
- Choose: "Collections" then "Heroes". Select hero: "Mei" and choose "Try" button
- In-game: choose Ally Hero to "None" and Enemy Hero to "None"
- Clone this repository (git clone https://github.com/pza007/game_bot.git)
- Install all packages, defined in requirements.txt
- Run the project while still in game
Before object detection:
After object detection:
Object detection starts with capturing frames of game's screen, using win32com packages. With the help of opencv package, I managed to detect position of the bot and minions. Please see below, the instructions that cover detecting the position of the minion, based on it's health-bar position:
- Convert RGB fame to HSV frame
- Create image mask, based on threshold color
- Transform mask (morphology)
- Detect contours of objects in the mask
- Execute filters:
- filter contours with 4 points (rectangle)
- filter contours with correct dimensions
- filter contours located outside skip areas
- filter contours with black border outside
Using combination of techniques mentioned above and 'matchTemplate' function, the following information can be extracted from the single frame:
'bot_dead': bool
'bot_pos_frame': {'health_bar': tuple[int, int], 'bounding_box': tuple[int, int], 'circle': tuple[int, int]}
'bot_pos_minimap': tuple[int, int]
'bot_health': tuple[int, int]
'bot_mana': tuple[int, int]
'cooldowns': {'Q': bool, 'W': bool, 'E': bool, 'R': bool, 'D': bool, 'well': bool}
'minions': {'blue': list[tuple[int, int]], 'red': list[tuple[int, int]]}
To control the bot, package pyautogui came in handy. Using it's functions I could command the movement of the bot by simulating mouse clicks on the game's screen. It also helped me to simulate more complex bot's actions, using key presses.
Another step in the project was to teach the bot how to efficiently play the game. For this purpose, I used keras, tensorflow and gym packages. I decided to use reinforced learning techniques. In this spirit, the bot (agent) will perform steps (actions) in the environment (game) and will get reward (experience) for it's steps. The results of the agent's learning will be transferred to model, that will update it's weights accordingly. Used parameters for the rl-machine-learning are presented below.
inputs:
# [0] float [0.0; 1.0] bot health = current / max
# [1] float [0.0; 1.0] bot mana = current / max
# [2] float [0.0; 1.0] distance on minimap between gate and bot's icon = distance/100
# [3] float [0.0; 1.0] distance on screen between bot and closest BLUE minion = 1/distance
# [4] float [0.0; 1.0] distance on screen between bot and closest RED minion = 1/distance
# [5..10] 6*int [0 or 1] 6*cooldown value
actions:
# 18*int [0 or 1] 18 possible actions:
# [0..10] 'move_up', 'move_down', 'move_right', 'move_left', 'move_up-right', 'move_down-right', 'move_up-left', 'move_down-left',
# [9..11] 'run_middle', 'collect_globes', 'hide_in_bushes',
# [12..14] 'basic_attack', 'q_attack', 'w_attack',
# [15..17] 'use_well', 'hide_behind_gate', 'escape_behind_gate', 'use_spell_d'
agent:
# policy = BoltzmannQPolicy()
# memory = SequentialMemory(limit=50000, window_length=1)
# agent = DQNAgent(...)
model:
# Flatten(input_shape=(1, 11))
# Dense(64, activation='relu')
# Dense(32, activation='relu')
# Dense(16, activation='relu')
# Dense(18, activation='linear')
# Flatten()
reward += amount of experience gained during step
reward += 5 # if action was successful
reward += 50 # if action involved attack
reward += 15 # if action involved gathering experience
reward += 5 # if bot is close to blue minion
reward += 10 # if bot is close to red minion
reward -= 2000 # if bot is dead
reward -= 100 # if action was not successful
reward -= 1000 # for incorrect actions, based on situation (bot played defensive when it has more than 10% of health)
The highest user-defined reward for the game was 1771. The highest game experience reward for the game was 1060 (episode 160 of training). The lower graph represents the number of all (blue line), successful (green line) and failed (red line) steps taken for each episode.
This agent was trained for 8000 steps. The end of session could occur because timeout (60sec) or bot's death. The rewards (upper graph, green line) were based only on the game experience, gained during each episode:reward += amount of experience gained during step
The highest game experience reward for the game was 1040 (episode 94 of training). The lower graph represents the number of all (blue line), successful (green line) and failed (red line) steps taken for each episode.
Alternative to machine learning is writing algorithm, using the game's knowledge. Code below chooses the action to be taken by bot, based on the following principles:
if no red minion is available:
1) 'hide_in_bushes' and wait
if red minions are available:
* 1,2 minions: 'basic_attack'
* 3>= minions: 'q_attack'
* 4>= minions: 'w_attack'
every 10 sec:
1) 'collect_globes'
2) 'run_middle' + 'collect_globes'
[finally] move behind the closest blue minion OR 'hide_in_bushes'
if bot mana < 5%:
1) 'use_spell_d'
2) 'use_well'
if bot health < 15%:
1) 'use_spell_d'
2) 'hide_behind_gate' + 'use_spell_d'
3) 'escape_behind_gate' + 'use_spell_d'
4) 'use_well'
The total game experience, gathered by the bot, looks similar for each of the three techniques. Unfortunately, some results were compromised by issue, observed at the end of the project. Not-picked experience globes were not cleared, after the game was finished. As a result, the bot from the next episode would collect the "experience from previous episode" and current episode. This situation was clearly observed, when the ai-driven bot collected all globes available for the episode and the value of rewards were similar for each episode. To overcome this issue, the waiting period would need to be implemented (around 1 min), after the game is finished. Unfortunately, this change will greatly increase the learning time of agents.
My personal winner is the ai-driven bot. It maximizes the damage done to the minions, the success rate of taken action is 97% and it uses escape or recovery techniques when at risk of dying.
The project is: no longer being worked on. The basic functionality was successfully implemented. There is plenty room for improvements to enrich the bot with further functionalities.
- Object Detection
- detect how the camera view distorts the environment when the bot is moving and compensate it (camera is not located directly above the bot but also shifted in x,y,z axis)
- detect experience globes and bigger "health/mana" globes
- detect fortification structure: gate, towers, forts
- Control of bot
- improve center of attack for W spell
- implement new actions to fight against structures and heroes
- Machine Learning
- improve type of model and(or) agent to get better results (higher value of game experience points)
- Other
- improve resetting game's state -> no experience globes are available at the start of new episode
Created by: [email protected]. Feel free to contact me!
MIT
Copyright to all images and texts belongs to: Heroes of the Storm™ ©2014 Blizzard Entertainment, Inc. All rights reserved. Heroes of the Storm is a trademark of Blizzard Entertainment, Inc.