DraftArtist

Example code for The Art of Drafting: A Team-Oriented Hero Recommendation System for Multiplayer Online Battle Arena Games

Authors: Zhengxing Chen, Truong-Huy D Nguyen, Yuyu Xu, Chris Amato, Seth Cooper, Yizhou Sun, Magy Seif El-Nasr

Point of Contact: Zhengxing Chen, [email protected]

Usage

The code shows our algorithm for hero recommendation in DOTA 2 Captain Mode.

Below is the command line to simulate drafts between two teams (p0 & p1) according to Algorithm 1 in the paper:

python3.6 experiment.py --num_matches=NUM_MATCHS --p0=STRATEGY0 --p1=STRATEGY1 

num_matches specifies how many drafts to simulate. The first team's first draft is always randomized in each match. The rule of randomized first pick is: regardless of the strategy adopted, the first hero is sampled following the probability distribution reflecting how frequently each hero is picked in our dataset.

Possible strategy strings are:

• random: randomly draft heroes
• hwr: always pick the hero not drafted yet and with the highest win rate
• mcts_maxiter_c: Monte Carlo Tree Search-based drafting, with maxiter iterations and exploration term c
• assocrule: association rule-based drafting

Examples

# Simulate 500 matches, with the first team adopting MCTS with 100 iterations and 0.03125 exploration strength, 
# and the second team adopting the association rule-based strategy:
python3.6 experiment.py --num_matches=500 --p0=mcts_100_0.03125 --p1=assocrule 
# Result
500 matches, p0 mcts_100_0.03125 vs. p1 assocrule. average time 2.09252, average p0 win rate 0.68968, std 0.15239 

# Simulate 500 matches, with the first team adopting the association rule-based strategy,
# and the second team adopting MCTS with 100 iterations and 0.03125 exploration strength:
python3.6 experiment.py --num_matches=500 --p0=assocrule --p1=mcts_100_0.03125 
# Result
500 matches, p0 assocrule vs. p1 mcts_100_0.03125. average time 2.12751, average p0 win rate 0.32866, std 0.15362 

[0.68968 + (1 - 0.32866)] / 2 = 0.68051, which is close to what we report (0.686) for UCT_100,0.03125 vs. AR in Table 5. (Results vary a little depending on the random seed being used)

Files

apriori relevant files for the association rule-based strategy.

• dota_lose_team_output.txt hero combinations appear frequently in the same losing team
• dota_oppo_team_output.txt hero combinations appear frequently in opposite teams
• dota_win_team_output.txt hero combinations appear frequently in the same winning team

input input file folder

• dota.pickle contains the processed dataset of 3 million matches from "Performance of machine learning algorithms in predicting game outcome from drafts in Dota 2" (Semenov et al., 2016).

models several models used in MCTS simulation

• hero_win_rates.pickle a dictionary recording each hero's win rate in our dataset. key: hero index, value: the win rate with value in (0, 1)
• NN_hiddenunit120_dota.pickle a neural network model trained by scikit-learn, used to predict win rate given a completed draft
• hero_freqs.pickle a dictionary recording each hero's selection frequency in our dataset. key: hero index, value: selection frequency normalized in (0, 1).

utils utility files

experiment.py the main file to start simulation

node.py implements MCTS-based simulation (UCT specifically)

captain_mode_draft.py implements the specific drafting rules of Captain Mode

player.py implements different team drafting strategies

Requirement

Python 3.6. Please also see requirements.txt