Please note that the visuals are designed for Sulfurous, but that can't show lists, so you can either have a beautiful, dysfunctional version, or an ugly, flashing, functional version. https://sulfurous.aau.at/#223801150 This is a variant of the Prisoner's dilemma in which the task is to gain as many points as possible, looking only at the possible results of if you and your opponent choose A (don't press button), or B (press button), and your opponent's behaviour responding to yours. You can create an AI to play this, or look at their scores and behaviour. Press space to see the statistics, and the AIs. The modes are: 1. Standard Prisoner's Dilemma, pitting together one of Always Cooperate, Always Defect, Tit for Tat, Grudger and Random. 2. Randomised Prisoner's Dilemma, all values for cooperation and defection are random, pitting together Random, Always A, Always B. Greedy, Minimax, Generous, Copycat, Beat Last, Average Best and Grudger. 3. Example of how nicer behaviours can be more evolutionarily advantageous: One Copycat, one Defector and 14 Cooperators. The AIs so far are (with colours, still in beta): 1. Random: Flashes random colours. Chooses A or B randomly. 2. Always A (or Always Cooperate): Always chooses A. 3. Always B (or Always Defect): Always chooses B. 4. Greedy: Always chooses the option with the highest potential payoff for itself. May find it similar to A or B, because it chooses one of these to keep doing. 5. Minimax: Chooses the option without the highest potential payoff for it's opponent. 6. Generous: Chooses the option with the highest potential payoff for it's opponent. 7. Tit for Tat (or Copy Cat): Chooses the same option as it's opponent's last option. 8. Beat Last: Chooses the option with the highest payoff, assuming it's opponent chooses the same thing again as it did last time. 9. Average Best: Chooses the option with the best average result. 10. Clever AI: Invented by me. If choosing the first option is always better than the second, it does so, and vice versa. Otherwise, it does the Average Best, but weights the options based on it's opponent's last decision. 11. Clever AI v2.0: Extension of Clever AI, also detects if it's opponent has always chosen A or B regardless of what it's done, bases result off that. 12. Clever Haxxor AI: Detects if playing against itself, then cooperates, or plays as the Clever AI v2.0. 13. Grudger: Works the same as Tit for Tat, but if you defect even once, it remembers, and always defects. 14. Nice for Tat (or Copy Kitten): Works the same as the Tit for Tat, but only 'Defects' if it's opponent has just defected twice. If we added a small miscommunication chance, but had a nice gene pool, then it would be better, but if we had a tougher gene pool full of Always Defects, it could get killed off more quickly. 15. Mean for Tat (or Copy Tiger): Works the same as Tit for Tat for the first two rounds, then only if it's opponent has cooperated twice, essentially a reverse Copy Kitten. If an opponent defects once, it will defect back twice. This makes it even worse at cooperating if there's a mistake. If playing against itself, and makes a mistake, it would only get 2 net points per subsequent round, as opposed to the Copycat's 4 or the Copykitten's. Still not as bad as the Grudger. :P ...And you can make your own AIs, if you remix the project, or even describe what it should do in the comments! Who can make the best AI?
You can see more about the original Prisoner's Dilemma at: https://en.wikipedia.org/wiki/Prisoner%27s_dilemma An interactive explanation of this, as well as the evolution of trust based on environment, which encourages you to always cooperate: http://ncase.me/trust/ Changelog: v2.1: -I'm preparing the visualiser! This will highlight the link between the two AIs playing in yellow, with everything else in white! That bit works, but I now need to make all points show their type with a coloured dot! Maybe a picture of a cat... -The Copytiger behaves as a Copycat for the first two rounds, not as Always Cooperate. This could have allowed for exploitation. v2.0: -One AI can be repeated multiple times in the programming, if you want to make your own mode to experiment with the programming! -6 more AIs to play with! -New game mode: Standard! This reduces the AIs, and keeps the points as they were in the original game. (I shifted it up by 4 points, so that it's easier to see.) -As long as their opponent cooperates, AIs like Tit for Tat cooperate blindly. But this doesn't always work as well as defecting. -You can now add a miscommunication chance! Notice how a small amount improves trust, while a large amount decreases it! v1.0: -Released. I was playing a really good text-based incremental game, Universal Paperclips, which is about being an AI tasked with manufacturing as many paperclips as possible, but whose programmers forgot to put a limit onto. One resource you could get was called Yomi. You could get Yomi by doing a version of the Prisoner's Dilemma. The goal is to get as many points as possible. The graph shows the AIs, and the points given to each. They can each choose A or B, which gives both AIs money based on the graph. The more complex AIs you can buy can analyse their opponent's past moves to determine what they should do next. You can play Universal Paperclips at: DecisionProblem.com/paperclips
