- 主题
- 回帖
- 0
- 精华
- 注册时间
- 2009-11-18
- 最后登录
- 1970-1-1
- 在线时间
- 小时
- 爆棚声望
-
- PGB
-
|
ZaZ 发表于 2021-12-31 10:03
Hi!
I'm the author of the tactic used in the experiments. I don't understand a single word of chi ...
Basic description.
1, we are a basketball manager game development team, the development process reference study a lot of FM settings, including the game engine.
We have learned a lot about the game from the bursting shed, and the purpose of this post is to give back to the community
2, in order to analyze the mechanism of FM's game engine, as well as the degree of science, we designed a system for measuring the degree of influence of each player's attributes on the final victory or defeat in FM.
3. fm-arena.com gives a preliminary test of the relationship between player attributes and wins and losses in FM2021 and FM2022, which is of some reference value and inspiration for our work
However, as it is only the non-professional work of amateurs, from the point of view of rigour, there are the following problems
1) It only gives the results of which attributes have a greater and lesser impact on winning and losing for all players, but in reality, the key attributes are obviously different for different positions, and the results of the test clearly show that only those attributes that are important for all positions will be more important, while those attributes that are important only for certain positions will be less important in the test. For example, its test results show that shooting has almost no effect on winning, while explosive power has a big effect on winning and losing.
2) Its test sample is insufficient, its test for each attribute was only carried out for about 900 matches simulated, but for a normally randomly distributed sequence, in general it needs to be randomized at 10,000 times before it converges relatively well to the mean.
(3) The effect of attributes on wins and losses is non-linear, and the test only deducts 4 points from the attribute to investigate whether it has an effect on wins and losses, but sometimes, just because 4 points have no effect does not mean that 8 points also have no effect, and it is also possible that 2 points have an effect that is close to 4 points.
4) There is a correlation between the impact of attributes on wins and losses, and its test of only changing one attribute at a time to test the impact on wins and losses can be interfered with by the correlation. For example, a breakthrough can lie on the ball to change direction and accelerate past a player, relying on physicality but not on discography, or it can rely on discography. If a player is good at both physicality and discography, then reducing his discography will not significantly affect his breakthrough effectiveness, as he can accelerate past people with a lie-back change of direction instead of discography.
5) Different tactics require different attributes for each position of the player, and it is not rigorous to talk about which attributes have an impact on winning or losing away from specific tactics.
4, we use artificial intelligence to conduct attribute importance research, artificial intelligence technology used a deep neural network similar to Go Alpahgo.
We fixed the tactic as the strongest tactic on fm-arena.com, ZaZ-Blue DM, for this experiment in order to analyse which attributes of players in each position have a greater impact on wins and losses under this tactic.
We built an artificial intelligence system and trained the AI to design the best attribute assignment for each position's performance with the same CA for each player in the whole team.
The approximate process was as follows: first, each attribute at each position was assigned the same value, then the neural network would try to change the value of certain attributes at certain positions (keeping CA constant) to see if the change had a positive or negative impact on wins and losses, and thus iterate over the multi-layer network to know what the AI thought was the best combination of 11 players.
We obtained the convergence results by training roughly 140 machines for 3 weeks and simulating 40 million games.
5. To verify whether the best 11-player attribute allocation given by the AI is really the best, we conducted another result validation test
Three groups of teams were designed, one with evenly distributed player attributes, another with the best 11-player attribute assignment given manually by experienced players based on their gaming experience, and the last group given by the AI. Put into a test league for 100,000 games, the AI's solution was significantly better than the human players' solution and far superior to the even distribution.
Conclusion.
For the ZaZ-Blue DM tactic, for each position we obtained the following levels of importance for the attributes (the values were normalized by 5 for ease of viewing, which should be sufficient accuracy for the game).
Winger
Defensive Midfielder
Defender(R L)
Defender(Centre)
Significance of use.
1, players can get the key attributes of each position according to the above table, 100 is the most critical attribute, while 1 is the least critical attribute, so as to guide the selection of materials.
2. The values in the above table can also be used as attribute weights to calculate the weighted average value of the attributes and then calculate the "Tactical True CA"
Tactical True CA = Weighted average value of attributes * 20 - 121
If the Tactical True CA is higher than the Player CA it means that the player is a good fit for the ZaZ-Blue DM tactics, the higher the Tactical True CA the better the fit.
This is used for player selection
For the convenience of future players playing in this way, the table we have given above is arranged according to the order in which the attributes are displayed in the player's interface, even if you don't write the program, you can also get the "Tactical True CA" in excel after quickly and manually entering the values of the attributes in three rows
3. For guidance on training, the attributes take up CA, so in order to train the players with the highest tactical true CA, we can train the attributes with the highest cost effectiveness
The cost effectiveness of each attribute for each position can be measured by the attribute Tactical True CA Weight / Attribute CA Weight, the higher the value, the higher the attribute will make the player's attribute Tactical True CA higher if it increases the same CA.
Main limitations.
The current experiments are costly and non-replicable, engine versions are changed, tactics are changed, and the AI needs to be retrained without migration learning, so our next phase will focus on migration learning, where the results obtained from iterative training in the case of a certain version of a certain tactic are used as the basis for a new engine and a new tactic, and rapid iterations are made to obtain results in a new environment. |
|
[复制链接]
发表于 2021-12-30 03:00
楼主
