In fact, this is a PPT speech a few months ago. On GDC 2018, Valve Games Engineer John McDonald introduced the resort's attempt to enhance CSGO users' gaming experience with deep learning. The reason why it is "improving the user experience" rather than "identifying the hanging behavior" is because V Club not only used the AI ​​technology in the anti-cheating mechanism, but also conducted thorough research on the psychology of ordinary players.
Ranked in the top 3 in the number of Steam games in January 2018
Hanging in CSGO
CSGO (Chinese name "Counter Strike: Global Offensive") is a first-person shooter team game launched by V Club in 2012. It is the fourth work of the CS series and has been popular with domestic and foreign players for six years. As an FPS game, it is almost impossible for players to open maliciously. Compared with the top PUBG in the above chart, CSGO has not been linked to the level of human anger.
However, players who hang up will be angry. When they are angry, they will complain and they will vote with their feet.
According to data disclosed by John McDonald, CSGO now has 12.4 million live students, with an average daily life of about 3 million, and an average team time of 1-2 seconds. For a game that has been launched for 6 years, it is not easy to maintain such data and stay on top of Steam. It is inseparable from the game environment created by the development team.
In 2016, V Club received overwhelming player complaints. Many people sued by sending e-mails, reddit postings, etc.: Your game has many hang-ups. Facing the player’s anger and the intractable situation of not being able to quickly investigate the hanging behavior, the V community is a bit embarrassed.
CSGO is a game developed by V Club. It comes with an official VAC system and can effectively ban openers. However, the popular games are always accompanied by economic benefits. Many people have developed a lot of loops to bypass VAC. , Or look for V clubs like to find the loopholes in the style after the fall, both took away the accounts of lucky players, but also ruined the game experience of normal game players.
To this day, many people are not forgetting to "pull too much" and criticize it when they are teasing that CSGO is "cool." The so-called deep love and responsibility. Their love for the game is real, but their abhorrence is also deeply bone marrow. The frustrating reality is that the cost of creating a CSGO plug-in is very low. If someone still keeps the plug-in of Half-Life 2, he will find that the link may actually continue to take effect in CSGO.
CSGO and deep learning
In 2015, CSGO's Overwatch was formally launched (don't think! I really didn't laugh). This is a supervision function. Players who are allowed to get the qualifications can judge whether they cheat or adjust the game parameters by watching Demo players who are suspected to cheat. . If players find that there is open behavior, they can submit the demo to the coordinator for further review, after which the official will follow up and feedback the results.
Overwatch itself can initially determine whether the player has opened or closed. The algorithm behind it is Naive Bayes. The system first finds a threshold for opening and then classifies the result. However, practice has proved that this system based on naive Bayes is too tolerant, it can only distinguish the obvious behavior of hanging, and then feedback most of the subtle Demo - slip the fish back to the real people.
For players, perhaps Overwatch is rather tasteless, but in the eyes of deep learning, this has become a rare source of data.
Model Introduction
The anti-cheating deep learning system developed by V Club is called VACnet. Because of the serial-oriented processing tasks, it was designed as a GRU RNN with a time step of 140, containing an input layer (256 neurons) and 4 hidden layers (256 neurons per layer). The activation function used by the network is ReLU, the loss function is binary cross entropy, and the optimization algorithm is Adagrad.
VACnet is trained in the data collected by Overwatch, which contains player-provided content and official collections. It is a large, high-quality data set. Although it is now also useful for deep learning to generate data, this approach makes it easy for neural networks to exploit loopholes, so developers are still using real game demos.
The input to the model is not a regular long vector, where they take 140 shots as a sequence and write down all the information for each video frame in order. These inputs are called (X,y), where X is a feature that we often say. It includes what weapon was used, what the shot was after, how far away (if hit) the target distance, and when the aiming angle was adjusted. At the end of training, the model needs to be able to fit the X data and repeatedly verify the quality of the study until the best results are output.
Inference and hardware
After training is completed, the model is used to infer (feed).
John McDonald compared this process to making beef burgers. When developers faced Overwatch, they had no ready-to-eat beef and puree in their hands, but instead they were constantly grazing and gaining weight. To make the data available for deep learning models, developers spent 95% of their time and energy cutting beef into beef—data, and then using the remaining 1% to crush beef—deep learning, 4% For making hamburgers - training models.
According to reports, V Club collects an average of 600,000 5v5 matches each day, which takes 4 minutes per game and a total of 2.4 million minutes CPU time (one day). However, only 1440 minutes a day, which means that in order to achieve the goal, they have to buy at least 1700 GPUs. At this time, the players dedicated to G fat worship began to play a role - they bought a total of 3456 CPUs. The following figure is the tip of the processor's iceberg. It contains 64 blade servers, each with 54 CPU cores and 128 GB of RAM, and is used only for model inference.
According to the anti-cheating logic established by the previous V community: players report - coordinator screening - player adjudication - open/unopened. Now that they have VACnet, a "VACnet report" is added to the report. It is understood that in all reports, the player's open-and-close verification rate is 15-30%, and the types are very rich; while the VACnet's opening and confirmation rate is as high as 85%-90%, but the type is relatively single.
Evaluation
As mentioned earlier, VACnet's effect is satisfactory. Despite its lack of flexibility and versatility, it has identified at least a large number of “crime cards†for openers. As shown in the figure below, when it was just added to the anti-cheating mechanism of the game, it slightly increased the number of system checks and hangs. After adding more training data, it again formed a sharp peak around 2018. And in terms of practical applications, it has not misjudged cheating incidents since VACnet was proposed so far. Of course, the last real person detection has also played an important role.
From the perspective of user feedback, this deep learning attempt has also achieved unprecedented success. Since the introduction of Overwatch and VACnet, the number of CSGO players complaining about reddit has drastically decreased, only 1% of the 2016 peak. This situation will be conducive to the survival of the old players and the arrival of new players, but also the V community loves to see.
But is VACnet really perfect? Actually not completely. In addition to the previous mention that it seems to identify only a few kinds of open hanging behavior, it also has a more serious drawback. During training, when the process of “Report - Coordinator Screening - Player Verdict - Open/Unopened†is completed, in order to ensure the learning effect of the model, the data determined by the real people will be fed into the model again to learn. .
This leads to a problem. If someone develops a hunt, it can self-delineate and lock the first few shots, and then return to normal after several shots. How does the system judge? If it is a real-life referee, he will certainly think: emm, this goods have problems, certainly opened. However, the machine does not know the basis for such judgments. It can only mark it as open, and then forcibly learns to open and hang from within the open time period and the unopened time period. This can lead to poor model performance.
To solve this problem, the current remediation method that V Company found is to add an automatic retraining function that can repair the loopholes caused by the heuristic algorithm. The second is to establish a brand-new model with no experience, which will expand the scope of open and hanging behavior detection, and increase the type of open and close that the system can recognize. The third is to apply VACnet to other steam games to pursue generalization.
summary
John McDonald believes that deep learning is an emerging technology that will reform traditional methods of anti-cheating. V Club has now used deep learning for anti-opening, anti-cheating, and Dota2 hero selection. Among them, the most concerned is Dota2's "big teacher."
For now, players do not need to worry about someone using deep learning technology to develop tools against VACnet, because the V community holds game data that others cannot imagine. For deep learning, the more data, the better the model performance. There is no doubt. The only thing to worry about is that this data set for training is collected from the game and someone may maliciously pollute the data. Although it seems that people can tell at a glance whether this person has opened or not, if the plug-in can do tricks, then it can fool VACnet and interfere with its training and learning.
Taking the whole speech together, we can conclude that VACnet is effective, but its anti-cheating ability is not as prominent as we thought. It still cannot completely eliminate plug-ins. If an anti-cheating system cannot be discerned even if the player has cheating, then its developer does not eliminate the determination of the plug-in, but only cares about the player's gaming experience.
In fact, this is where VACnet succeeds. It immediately reduces the number of complaint posts, stabilizes the player's heart, and positions CSGO in the top three positions in the player list. When players again talk about the open and hanging behavior of the game, they will also be happy to discuss the V-community's efforts in anti-cheating, and praise the actual effect of deep learning.
For an ordinary player, is the result like happiness?
With more than 15+ yrs rich MFG experience, you can definitely trust in and cooperate with.
Provide you with the supply of Personal Protective Equipment. to help you safely get back to your daily routine.
Our products include pulse Oximeter Finger, Forehead Thermometer, Automatic foam soap dispenser, etc.
Our strict quality control protocol thoroughly vets every aspect of production, storage, and shipments all the way way to our end customers.
infrared thermometers wholesale, forehead thermometer wholesale,wholesale thermometer suppliers
TOPNOTCH INTERNATIONAL GROUP LIMITED , https://www.itopnoobluetoothes.com