Yes, I think Tibia is using sever-sided heuristics (logistic regression or even ANNs) to detect bots. XenoBot stands up to this in many ways: timing randomization (healing and such) and walking randomization (nodes and weak nodes) are two of the major ways. Other things that make XenoBot safer is the prevention of absurdly fast actions (no other bots do this), lack of an after-death relogger (I often state my reasons for this), and a very human-like spell shooter (sometimes casting when exhausted, sometimes shooting a less-optimal GFB, sometimes trying to shoot through walls; all of these things are never done by other bots).Originally Posted by Jesseh
Little variations in all of these actions steeply degrades the integrity of ML-base detection algorithms. This is because the ML-based heuristics used in bot detection are designed for one purpose: classification. Typically, for this implementation, there are two classifications: bots and non-bots. Because XenoBot behaves differently than other bots, it adds a lot of noise to the "bot" classification during training and makes it hard to select from prime components, even when using certain types of feature selection such as PCA or random forest. Not to mention, XenoBot is much smaller (user-wise) than WindBot, so CipSoft would likely optimize their training set to improve their classification of Windbot, not XenoBot.
Of course, there's an alternative. Their classification engine might have many outputs: non-bot, XenoBot, WindBot, iBot, and BBot. In this case, however, the problem becomes much harder. For each new classification, you double the amount of nodes and synapses in the network (I'm assuming they use an ANN, here, as the features wouldn't fit well with logistic regression). That means: training time is exponentially harder, as the network must be fine-tuned using multiple rounds of feedback; processing takes more time, as the network has far more connections to evaluate; and detection has a higher false-positive rate, as you increase the chance of mis-classification for every class you have.
Of course, this is all speculative and highly summarized to be at least somewhat digestible for an audience that is not well-versed in machine learning or computational neuroscience. In the general sense, though, the assumptions herein are correct, and I stand by these reasons for why XenoBot is better at resisting detection than other bots (namely: WindBot).
