Edit: fixed bug with floating-point value overwritten as an integer :/. In this scenario, you give an AI $1,000 for gambling. Can it make a profit, or will it bust easily? A classic proof that even a 50/50 chance isn't fair. In fact, almost any chance of 50% or below always results in high net loss followed by an eventual bust, no matter how risky you are. The green line of the graph is the AI's money. Bigger peaks and drops are indicative of a larger bet. The blue line of the graph is the "confidence", or how sure the AI is that it will win this turn. As the number of consecutive wins goes up, the confidence goes down; as the number of consecutive losses goes up, the confidence goes up. This is due to the Gambler's fallacy: that the chance of a win after a bunch of losses increases, and vice versa. Given that no matter how confident the AI is, it is still at the same risk of being wrong, and thus cannot truly predict the outcome. The red line is the "successfulness" of predictions. The higher it is, the better the AI is at predicting an outcome. If it stays near the median, it means the AI is having difficulty predicting consistently. Risk Factor: Risk factor determines how much the AI is willing to bet. A risk factor of 0.01 means the AI will bet at most 1% of its money, while a risk factor of 1 is basically double or nothing. Experimentation Notes: Having a near-100% chance of win means the AI will make the mistake of predicting a loss even though one may never come for a while. This is another flaw of the Gambler's fallacy at work. Setting the chance of win very low will create a long descending line as the AI knows it cannot bet more than the minimum, which means its money pool is slowly drained. As such, these risks are not worth gambling on. Interestingly, a chance just below 50% is actually more reliable than at 50%. Sources: https://en.wikipedia.org/wiki/Gambler%27s_fallacy
