Title: CoG-2023 Multi-Agent Google Research Football Competition 

Organizers: 
Haifeng Zhang, Institute of Automation, Chinese Academy of Sciences, China

Yahui Cui, Institute of Automation, Chinese Academy of Sciences

China Yan Song, Institute of Automation, Chinese Academy of Sciences, China 

Website: http://www.jidiai.cn/cog_2022/ 

Description: 
The competition will be named CoG-2023 Multi-Agent Google Research Football Competition as it plans to use the multi-agent scenarios of the Google Research Football (GRF) simulators [1] as the testbed. The GRF simulator bears great similarity to well-known football games like FIFA and REAL Football, vividly modelling the dynamics of the ball and player movement as well as their interaction such as passing, shooting, etc. The competition focuses on the full game scenario built on GRF simulator where two teams of players compete for goals on the pitch (as shown in Fig. 1). The competition will be held on Jidi platform which offers online evaluation of submitted policies on various simulated environments and holds Kaggle-like customized competitions. Regarding the scale of the game, the GRF competition has two tracks as two separate contests, one easy track using the 5vs5 multi agent full-game scenario and one hard track using the 11vs11 multi-agent full game scenario. 

On these scenarios, the duration of the game is 3000 discrete time-steps corresponding to 90 minutes in a real football match. The observation for each player contains information of both the teammates and the opponents, as well as the state of the ball. Based on the provided information, each individual player can either move or make specific actions such as ball-passing or ball-shooting. The goal of each team is to control all of its players (5 or 11), organize team offense and defense and win the game by scoring more than the other team. Expectedly, a strong participant is required to demonstrate sophisticated football skills and counter various opponentsʼ tactics in such complex competitive tasks. 

Previous GRF Competitions: 
GRF has already been chosen previously as a testbed for AI competitions. For instance, the Single-Agent RL (SARL) Google Research Football 2020 Kaggle Competition [2] was held by the famous Manchester City F.C., attracting over 1,000 team participation on the single-agent 11v11 full-game scenario. In this task, one side of the team only needs to control the single designated player and all of its teammates are controlled by the built-in AI. In year 2022, Jidi held the MARL CoG 2022 GRF tracks [3] on multi-agent RL (MARL) settings (5v5 and 11v11 full-game scenarios) with an increase in the number of controllable players.This was more difficult than the single-agent settings and had attracted over 100 team participation (see Fig.2). 

Jidi plans to continue the competition on GRF in year 2023 after discovering interesting tactics from MARL CoG 2022 GRF tracks. On the 5v5 track particularly, many top participants have learned a high-pass glitch existing in the goal-keeperʼs built-in logic and taken advantage of such loopholes to form an extremely strong tactics. On the 11v11 track, some collaborative behavior between players have emerged and a game clip can be found in [4]. To further encourage multi-agent learning research, we will hold two tracks on 5v5 and 11v11 full-game scenarios respectively: 

1. 5vs5 track: http://www.jidiai.cn/env_detail?envid=71; 

2. 11vs11 track: http://www.jidiai.cn/env_detail?envid=34;

Evaluation Details: 
Each team of the competition consists of 1-3 members and each track is scheduled for two warm-up rounds and two main rounds. Only the results of the main rounds are added up to give the final scores and the team with the highest score wins the track. Both the warm-up and the main rounds adopt the Swiss Round double competition system [5] and the ranking of each team is determined through (4⌈ logN⌉+1) rounds of matches (N is the number of teams). The double competition system means that the two submissions will be evaluated under the same environment random seeds for two rounds and switching sides when each round finishes. The initial score for each team is at 1000 and are updated through ELO algorithm. 

Tracks: To further encourage multi-agent learning research, we will hold two tracks on 5v5 and 11v11 full-game scenarios respectively: 

1. 5vs5 track: http://www.jidiai.cn/env_detail?envid=71; 

2. 11vs11 track: http://www.jidiai.cn/env_detail?envid=34;