forked from google-research/football
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathCHANGELOG
More file actions
56 lines (48 loc) · 3.38 KB
/
Copy pathCHANGELOG
File metadata and controls
56 lines (48 loc) · 3.38 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
Numbering scheme:
Each environment release number is of the form vX.Y. X is the major version
number, Y is the minor version number. Experiment results within the same X
should not change, as modifications made to the environment are either
new features or backward compatible bug fixes. We will maintain vX branches
pointing at the most recent vX.Y.
v1.5
- Fix to numerical issue resulting in asserts in long runs (https://github.com/google-research/football/issues/78).
- Eliminate dependense on Glee (https://github.com/google-research/football/issues/77)
- Created google-research-football@google.com email for direct contact with the environment team.
v1.4
- Added implementation of architecture 'gfootball_impala_cnn' used in the paper.
- Added possibility of loading PPO checkpoints as players, added example checkpoints (for levels 11_vs_11_easy_stochastic and academy_run_to_score_with_keeper).
- Removed TensorFlow dependency when running only the environment (but for training OpenAI baselines it's still needed)
v1.3
- Fix to pixel representation (https://github.com/google-research/football/issues/54,56,57).
v1.2
- Reduction of memory usage by the environment (https://github.com/google-research/football/issues/47).
- Reduction of memory usage during compilation (https://github.com/google-research/football/issues/49).
- Elimination of --no-pie compilation flag (https://github.com/google-research/football/issues/53).
- Fix for broken wrappers.py (https://github.com/google-research/football/issues/54).
v1.1
- Add support for running multiple environment instances in a single process.
- Replaced left_players and right_players with a single players flag. This allows for controling players in both teams by a single agent (useful for self-play).
- Make ball_owned_player and ball_owned_team consistent (fix for https://github.com/google-research/football/issues/45).
- Removed paper.pdf and references to it. The paper is now at https://arxiv.org/abs/1907.11180.
v1.0 (released July 19th, 2019)
- added support for multi agent training with run_multiagent_rllib.py example.
- added enable_sides_swap parameter to create_environment, which enables random swapping of team sides for each episode. This is useful for self-play training.
- over 3x speedup to the game engine.
- renamed home/away teams to left/right.
- eliminated implicit observation and action rotation for the right team. `is_active_left` denotes controlling left team.
- config is not passed anymore to the reset method.
- minor game engine bug fixes and code cleanup.
- eliminated chdir calls from the environment, so that working dir does not change.
- backfilled missing values in simple115 representation for the case when training with less than 22 players.
- updated game engine to use Boost.Signals2.
- updated environment to the new OpenAI Gym API.
v0.3 (released June 14th, 2019)
- change action set to include sliding, dribbling and keeper rush.
- cleanup in action set & sticky actions on python side.
- change to env: now it release non-sticky actions after a single environment step.
- game engine change: the goalkeeper now move backwards when the ball if bound.
for the goal and is flying high.
- game engine change: call randomize with seed right before match creation.
- update documentation with frequent problems & mailing list.
- updated paper.pdf with citation to RoboCup and new action set.
v0.2 - initial release (released June 7th, 2019)