CS:GO strategies, analysed: A look at how Bayes Esports analyses data using AI
From our backer: Bayes Esports
In this piece, Gustav Geissler, Data Scientist at Bayes Esports, describes how Bayes analyzes CS: GO strategies using AI.
Choosing and executing a strategy can make or break your game. When two equally strong teams face each other, Teams Strategies will affect the outcome much more than the skill of the individual players. Strategic thinking becomes even more important when facing a stronger opponent. While you can’t play them mechanically, you can still win if you’re smart. Playing (and winning!) At a professional level therefore requires a thorough analysis of your own strategy and that of your opponent before the game. This is old news for chess, soccer, or any other competitive sport, and it applies to CS: GO as well.
The status of the strategy description
In general, a strategy is a way of addressing a problem that leads to your preferred outcome. In the context of CS: GO, we see winning a round as the goal a team strives for and the combination of player path, weapon choice and grenade use as the strategy they use to achieve it. In this article, we’ll only focus on the spatial component – the routes and locations players choose at the start of each round. For the sake of simplicity, we’ll call this the strategy a team plays.
Position analysis like this has long been done in a categorical manner for most classic sports (think opening moves like the Queen’s Gambit in chess or formations like 4-4-2 in Football). In comparison, strategies in CS: GO are still either described in a descriptive way (‘Bomb + 1 A short, rest A long, smoke middle box ‘) or by name (‘3 + 2 B-Split’), which leave much room for interpretation. The way in which the same strategy is invoked can also vary significantly from source to source. At Bayes Esports, we believe the time has come to professionalize this part of the game.
Convert strategies into numbers
Think about your average CS: GO card. It has two bomb locations and several ways to get to them from the terrorist spawn area (In CS: GO, the terrorist side is the attacking side; the goal here is to plant the bomb while counter-terrorists aim to prevent or defuse a bomb installation). The attacking team must therefore decide which bomb location to target, who will carry the bomb and, ultimately, where each individual player will go. Just as the choice of grenades and weapons play a big role, the players’ movement paths are perhaps at least as important to the team’s success in winning a round.
At the beginning of our analysis we could observe all strategies ever played on a card and assign numbers to them. This would be the simplest approach for integrating strategies into a system, but would have almost no value for use in the game due to its poor description quality. Also, the sheer number of strategies you need to memorize in order to use this approach makes it practically useless.
Of course we need a new system that will make the lives of players and coaches easier. Not one that works without numbers, but one where the numbers are more descriptive and don’t require a lot of memorization. Bayes Esports has developed such a system that gives each strategy a number. With relatively little additional information, that number alone tells you what the strategy looks like.
Strategy as the sum of paths
When we look at strategy, we usually look at how the team is moving as a whole. But how important is the individual player for the whole? What if only one player takes a slightly different path – is it still the same strategy? If not, how many players will have to change their path to make a new one? And what happens to a strategy if all players follow the same path, but the bomb is carried by another player the next time the strategy is executed?
From our point of view, all of these details are important. They should therefore be clarified by describing a strategy. So instead of looking at the teams’ movements as a whole, let’s break them down to see what individual movements it actually consists of. Then a strategy is nothing more than the combination of the paths of all players. Add an extra path for bomb movement (as it can be dropped by one player and picked up by another player) and you’ve got a pretty accurate description of what’s going on in a CS: GO round.
In fact, by analyzing tens of thousands of rounds played by professional CS: GO players, we have found that almost all (> 98%) strategies can be broken down into a combination of fewer than ten different ways that the players will perceive take according to the card. That sounds like a small number, but the number of possible strategies they can be combined into is large. For example, let’s say there are four ways to attack a bomb site and all players randomly choose one of them. That alone (and taking into account the various possible bomb carriers) gives us a whopping 280 different possible strategic approaches for just this one location!
Our system can then describe each of these strategies with just six numbers: one for each player’s movement path and one for the bomb.
Our approach to CS: GO strategies
In order to recognize the strategy being played, we trained a convolutional neural network to recognize the path of each individual player. The clustering process divides the paths found in thousands of rounds into about seven to ten clusters, i.e. different paths, and assigns the respective number to them. The strategy of a particular round can then be described by the number of paths taken. In this way any historical and even running game can be analyzed.
Let’s use Dust2 to take a closer look at how this works. There are eight possible routes a T-side player can take at the start of a round. We randomly assign them the numbers 1 to 8.
The illustration below shows an aerial photo of a Dust2 cartridge on the left. The paths of all terrorists are shown in different colors. We see that the team is divided into two groups, each taking a different path. Even if the individual players in each group move a little differently, our neural network can recognize which path they are taking. We have assigned the numbers 2 and 4 to these paths. In this example, three players take path number 2 and two players take path number 4. Suppose the bomb path matches the path of a player who chooses path 4 in this example. Then we can describe this strategy as 4-22244, with bomb movement being the first number by convention. Since we are not tracking individual players, the other path numbers are sorted in ascending order.
With our neural network, the process of extracting and labeling these paths can be fully automated. The results of the detection are displayed on the right in the picture above.
We can describe any strategy played on Dust2 in a similar way, regardless of which players choose. Once you’ve memorized the numbers of the various paths, analyzing T-Side strategies on the map becomes a breeze.
CS: GO strategies: Statistics on Dust2
After doing this for a large number of professional games, statistical analysis of CS: GO games becomes surprisingly simple. Here are examples of statistical values that might be of interest to players, coaches and spectators alike, computed by Bayes Esports. These numbers only apply to Dust2’s terrorist side, which has used thousands of rounds of professional games (including mostly the top 30 teams in the world but other teams close to them as well) over the past 18 months.
22% of the rounds are played with the three most common strategies:
- A long rush (11% – strategy 3-33333)
- A long rush with a detour at B (7% – strategy 3-33334)
- Brush (4% – strategy 1-11111)
80% of all Dust2 rounds can be described with 160 different strategies (combinations of individual paths):
- Bomb site A is attacked in 76% of all rounds
- Pistol rounds are played at high speed in 31% of the rounds (A: 22%, B: 9%)
- In 42% of the rounds, four or five players go the same way
- The bomb carrier takes one route alone in 17% of all laps
But we can go even deeper and compare strategies with the economy of a team, the probability of winning a round and other measures. We can look at individual team and player preferences. And of course we can add CT behavior (in addition to the T-side behavior) to get a full picture of how two teams are playing against each other.
This new approach to CS: GO strategies makes it easy to systematize and compare them. In contrast to individually named strategies, our numbering system is not open to different interpretations and does not require that we come up with any unique identifiers and then memorize them. It all happens automatically. Once the strategies are identified, it opens a door to new types of statistical analysis of CS: GO games, be it for a deeper understanding of the game mechanics, preparation for an upcoming game or more detailed in-round information for esports betting. Watch and comment live.
See more new articles in category: Features