forked from google-research/football
-
Notifications
You must be signed in to change notification settings - Fork 0
/
CHANGELOG
112 lines (92 loc) · 5.79 KB
/
CHANGELOG
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
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
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.
v2.6
- Address minor issues reported by Kaggle contestants.
v2.5
- Make 2D videos more readable and color-blind friendly.
- Do not allow agents to block the game by exploiting kick-off.
v2.4
- Support for configurable rendering resolution.
- Kaggle competition links.
v2.3
- Fix to scenario's second_half feature.
v2.2
- Minor changes to the information displayed in the video replays.
- Changes to Kaggle competition scenario.
- Minor code cleanups.
v2.1
- Introduction of `builtin_ai` action, which lets agent dynamically give control
of the player to the native game's built-in AI. It can be useful when training
multiagent setup incrementally (for example start by controlling all 11 players,
for 10 of them generate `builtin_ai` action. Over time stop using `builtin_ai`
action for an increasing number of players). Introduction of `v2` action set
which is an extension of the `default` action set with `builtin_ai` action.
- Improved documentation for using Docker image, support for rendering with Docker.
- Fix to observation shape returned by wrappers.
v2.0.7
- A number of documentation updates.
- `info` returned by step() API call returns pointers to recorded dumps.
v2.0.6
- Small fixes to getState/setState.
- Memory usage reduction when saving replay videos is enabled.
v2.0.5
- Support compilation on OSX.
v2.0.4
- Added Colab support (see readme file).
- Some minor bug fixes (including https://github.com/google-research/football/issues/110)
v2.0 - this release is not backward-compatible due to making game fully symmetric (training agents can behave differently).
- Game is fully symmetric now (playing left to right and right to left is equivalent). All observations are modified inside of the
environment and from agent's perspective the game is always presented as playing from left to right (no need to train from both
sides).
- Introduction of get_state() and set_state() API methods, which allow for saving and loading state of the entire environment.
Using these methods it is possible to move back in time and investigate possible outcomes of different actions taken
at exactly the same non-initial state of the episode.
- Introduction of GYM-like render() method to the environment API which enables rendering dynamically. It is not needed anymore to
pro-actively start environment with --render flag in order to see rendered episodes at any time.
- By default team losing the goal starts the game with the ball.
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 [email protected] 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)