Simulation of the prisoner's dilemma in Python.

Inspired by the Veritasium's video and a browser game by Nicky Case called The Evolution of Trust

Instantiate the simulation object as specified in the constructor or call the suite function to run a series of simulations.

There are 2 modes of simulation:

• Round-robin, where all strategies play each other exactly once
• Evolution, where all strategies play a round-robin tournament and, afterwards, the bottom 10% of strategies get eliminated and replaced by the top 10%.

As a twist, I introduced a strategy called machine learning, which incorporates a q-learning algorithm

The list of strategies that have been implemented is as follows:

• always cooperate - always cooperates
• always defect - always defects
• tit-for-tat - picks what opponent picked in the previous turn
• grudger - cooperates until cheated
• pick random - 50/50
• suspicious tit-for-tat - tit-for-tat but starts with defecting
• tit-for-two-tats - defects if opponent defected twice in a row
• two-tits-for-tat - retaliates twice after being cheated once
• pavlov - if the opponent's move was the same in the last turn, cooperates, otherwise - defects
• detective - cooperate, defect, cooperate, cooperate, then always cooperate if not cheated, always defect otherwise
• simpleton - if the last move cause a point gain, repeat it, otherwise change it
• coop-75 - cooperate with .75 probability
• retaliate-75 - cooperate, unless cheated - then retaliate with .75 probability
• machine learning - utilizes q-learning. The model is instantiated from scratch every time the simulation object is created, so the learning process should happen through the suite function rather than reinstantiating the simulation

There are duel and duel_all functions that enable to just battle selected strategies. They can be used to analyze the machine learning model after the simulation(s).

I have analyzed mostly two sets of players: exhaustive - having each strategy once, hostile - focusing on strategies that want to cheat or backstab sneakily

My foundings:

• In exhaustive tournament, the machine learning strategy mostly defaulted to something akin to the tit-for-tat strategy, in more extreme situation switching to the grudger
• Sometimes, the model would learn to cheat on the detective strategy in turn 7 to gain back what was lost in turn 2, which I found to be interesting
• In hostile environment, the model seemed to default to always defect

There are some .txt files that contain duel_all results of the machine learning strategy after some simulations.