The Shattered Web update has been a total blessing for the CS:GO community as this update finally brings an Operation into the game after years of stagnancy. We all got busy hunting for stars and completing the weekly missions when the update launched, but there was a small change we all might have ignored, that was the proficiency with which the bots attacked us in the exclusive Guardian missions. The bots surprisingly seemed very well-coordinated and even used smokes and flashbangs at times, similar to real human players and their behaviour.
The improvement in the CS:GO AI has been exponential ever since Valve announced that they will be improving the in-game AI in their September 2019 update. The update had introduced decision making for bots based on behaviour trees as an early framework which Valve hoped to improve on. The most notable change to the AI in CS:GO is possibly observed at random times in Valve Deathmatch when a bot can just one tap you from nowhere before you can even react or just land an impossible flick shot in your face. Each individual bot in CS:GO takes in information like sounds and sightings of the enemy and processes them to act on and carry out actions which are limited to their logic code. They even have their own form of memory (short term and long term retention capability although both lasts only for mere seconds). The bots most notably seem to prioritize objectives over most things as per their code. Currently, we can say that Valve has done a lot of work to make their bots imitate human-like aim, but what really impressed us is how they have designed the bots in Guardian Missions allowing them to even use utilities to their advantage just like humans play.
Sadly, the decision making of the Guardian AI isn’t fully self-learnt or automated as the developers have uploaded specific logic codes for each mission which explains why their executes although pretty good, were fairly predictable, to begin with. In the end, the behaviour tree-based decision making as implemented by Valve isn’t complex enough to fully automate the entire decision making of a bot like a neural network-based decision-making system, where almost the entire thought process of a bot is automated based on the information received and it becomes a self-learning process where the bot can come up with some really complex strategies which the players are totally not prepared for based on how the game is going. In the current state, the in-game AI is very limited in what it can do but it has come a long way in terms of improvement and growth.
Although there is much more to be wished for, it is still good to know that Valve is doing their best to improve their in-game AI and trying to implement more human-like features to the bots to make them even more competitive.