Where do I throw my nades?

A data driven analysis of strategic grenade usage in Counter Strike: Global Offensive

By Kyran Adams

Counter Strike: Global Offensive (CS: GO) is a team-based strategic shooter in the Counter Strike series, which has nearly 20 years of competitive history and tournaments now often offering prize pools of more than $1,000,000.

One thing that has always drawn me to this game is the strategic importance of utility — smoke grenades (smokes), flashbangs (flashes), and high explosives (HEs). A well-thrown smoke can give your team cover to advance into a new position or block eyesight from an enemy, helping your team win the round.

Keep scrolling. Using data from this kaggle dataset, we will analyze thousands of rounds of CS: GO to determine the best positions to use utility. Pictured on the right is an aggregation of all of the smokes thrown by Counter Terrorists across thousands of games.

CS: GO

First, I will give a small explanation of how the game works.

A round is between two teams of five players. In each round, one team, the Terrorists (T), tries to plant the bomb on one of the bombsites and defend it until it detonates. The other team, the Counter Terrorists (CT), attempts to stop the bomb plant or defuse it once it's planted.

On the right, we have grenades through by Ts (red) and grenades thrown by CTs (blue) in a single round. The CTs won this round.

Now, to do some more in depth analysis, we're going to combine all of the utility used from each team by type. Here on the right are smokes thrown by CTs in every round.

In order to get some idea of how location of utility relates with effectiveness, we could color grenades based on the result of the round, which we we will do next.

Effectiveness of Smokes

Here, all of the smoke grenades are colored red if that round was won by the thrower and blue if it was lost. However, the patterns are a little unclear and it's hard to figure out which locations are the best locations to throw grenades.

In fact, we're not just interested in the the impact of individual points, we're more interested in grouping smokes together and determining the average impact for that group. To do that, we can use a clustering algorithm.

Clusters

One clustering algorithm, DBSCAN, creates clusters that correspond to high density areas of points and leave lower density areas outliers. We can use this to color the smokes on the right based on their cluster.

We can then run logistic regression using these cluster counts as predictors. We are basically asking the question: What is the average change in the likelihood of winning from throwing one additional grenade in any particular cluster?

Effectiveness of Counter Terrorist Smoke Clusters

Here are the results, where the average effect of a cluster is plotted as the color of that cluster. The more blue a cluster is, the more it helps CTs. The more red, the more it helps Ts. The color of the outliers (a non-clustered smoke) is based on the average effect of all of the outliers, which has a large variance.

From this graph we can see that the most effective smokes are mid-connector, t ramp (especially deep ramp), and palace. In fact, in relation to the other types of nades, CT smokes are very effective.

Effectiveness of Terrorist Smoke Clusters

Terrorist smokes are more variable in their effectiveness. Some are extremely effective, and some actually help the CTs (this could be because they warn the CTs of an incoming push). The most effective smokes are top mid, stairs, front halls, A site, kitchen window, and jungle. The most ineffective smokes are jungle/connector, jungle, lower mid, palace, top mid. Sandwich is the only statistically significantly bad for the Ts, which makes sense, as it is not an intentional smoke, and usually signifies a missed stairs smoke.

One interesting result is that some clusters are very close together, but have very different effects. One example is that a deeper top mid smoke onto catwalk is much more effective than a shallow one. Similarly, missing the jungle smoke by even a bit reduces the effectiveness considerably.

Effectiveness of Counter Terrorist Flash Clusters

It seems that for CTs, flashing to take underpass is very strong. This makes sense as it is very difficult to defend against this flash, and strategically, CT taking underpass control cuts off half the map for Ts. It also seems cluster connector, T ramp, and outside T ramp are very good. Mid window flash is not very good. It is interesting that CT flashes seem more effective than terrorist flashes in winning the round, and more statistically significant.

Effectiveness of Terrorist Flash Clusters

For terrorists, the strongest flashes were outside apartments and underpass. It is interesting that top mid over ramp are not very effective at all, despite being very common. It is very interesting that outside b apartments, one of the most common flashes for attacking B, is actually negatively impactful. This could be because it alerts the enemy of the attack rather than surprises them.

Effectiveness of Counter Terrorist HE Clusters

It seems that CT HE's are very effective, especially in top of mid, palace, and underpass.

Effectiveness of Terrorist HE Clusters

Many terrorist HE's also are very effective. Mid window, connector and jungle are incredibly effective. HE's are also interesting in that they are quite effective also when not in any cluster at all (the -1 cluster).

Effectiveness of Terrorist Groups of Smokes

I performed logistic regression on just terrorist smokes, but added interaction terms between groups of smokes that I thought might be significant, such as CT + stairs + connector/jungle for taking A.

I found that out of these groups, none had a significant increase in chance of terrorist wins, and most actually decreased the chances that terrorists would win. My hypothesis is that often at lower ranked play, players throw smoke groups simply because they see professional players doing it, and not actually for a strategic advantage. However, when players throw individual smokes, they do it more strategically. .

Effectiveness of Utility as a Whole

Finally, I aggregated all of the data together and ran logistic regression to find the average effect of utility by its team and type.

Interestingly, terrorist flashes are the weakest out of all types of utility, with almost a negligible effect on round wins! HEs are also surprisingly strong.

Takeaways: How should I use my utility?

Through this analysis we found many interesting results impacting strategy (at least at the nonprofessional levels from which this data was collected). Here are some takeaways to guide utility usage on de_mirage.

• As a T, one should be careful using flashes without thought. The most commonly thrown flashes are actually very ineffective - top mid, T ramp, and over B apartments. In fact, common flashes (those in clusters) are not much better on average than the rest.
• As a counter terrorist, almost all common flashes are good.
• HEs are much more effective than I previously believed, for both teams. CT HEs are effective in clusters and outside of clusters. T HEs are effective in some clusters and out of clusters. Some T clusters are ineffective and should be avoided.
• Smokes are, as expected, the most effective type of utility. They have a strongest net effect from Ts, even though CT smokes have a more consistent effect across the different clusters.

The biggest takeaway from this is to be intentional with utility. Many of the ineffective uses of utility that I found align with common patterns at lower levels, which I hypothesize are simply done out of habit rather than strategy.

Thanks for reading!

Questions or thoughts? Tweet me at @Kyran_Adams or email me at kyran.park.adams@gmail.com.