pokers 0.1.2

Embarrassingly simple No Limit Texas Holdem environment for RL

Pokers

Embarrassingly simple no limit texas holdem environment for RL.

Why another poker environment?

Poker is a incredibly deep game with very simple rules, so why are all the environments so overly complex? Heck, someone could say that you need to publish a paper before building one (looking at you RLCard 👀). Pokers way is to discard the agent environment cycle and all that stuff, just the good old new_state = state + action model. Through its simplicity pokers tries to be flexible and easily integrable into any framework.

Why not to use pokers

Pokers is a side project inside another side project. This means that it is guaranteed to have bugs, which is not very nice for a RL environment. We have done our best to minimize the errors, testing it against the 10k hands pluribus logs. However, this doesn't cover some areas of the state space, so if you need a more reliable environment RLCard is a better option.

Installation

Pokers can be installed directly from pypi.

pip install pokers

Usage

Just create the initial state and act over it. Easy peasy.

import pokers as pkrs

agents = [agent0, agent1, agent2, agent3, agent4, agent5] # Build the agents however you want
initial_state = pkrs.State.from_seed(n_players=len(agents), button=0, sb=0.5, bb=1.0, stake=100.0, seed=1234)
trace = [initial_state]

while not trace[-1].final_state:
    state = trace[-1]
    action = agents[state.current_player].choose_action(trace)
    new_state = state.apply_action(action)
    trace.append(new_state)

The initial state can also be declared with a fixed deck with State.from_deck().

Curious about what info a state contains? Just go to pokers.pyi and see it yourself, I bet there's all you need.

As a bonus you can print the entire hand as text. Who wants GUIs anyway?

print(pkrs.visualize_trace(trace))

Error handling

There are two possible types of erroneous states: when an illegal action is performed and when a player bets more chips than he has available. These cases are represented by the enum StateStatus with the values IllegalAction and HighBet, the value Ok is used for correct states. This information is stored in the field status of the state so you can filter them.

Every erroneous state is also final. So applying an action over it will return the same exact state.

Parallel actions

If you have a bunch of independent states and want to perform multiple actions in parallel you can easily trick the GIL with parallel_apply_action().

import pokers as pkrs

agents = [agent0, agent1, agent2, agent3, agent4, agent5]
states = [pkrs.State.from_seed(n_players=len(agents), button=0, sb=0.5, bb=1.0, stake=100.0, seed=seed) for seed in range(10)]

while not all([s.final_state for s in states]):
    actions = [agents[s.current_player].choose_action(s) for s in states]
    states = pkrs.parallel_apply_action(states, actions)

Since final states do not change when an action is performed, you can safely wait for all hands in the batch to end.

Alternatives

To our knowledge these are some other poker environments that you would want to consider.

• RLCard: Great RL environment for multiple card games.
• neuron_poker: OpenAI gym for texas holdem.
• pgx: Pretty cool project with jax-native game simulators. Sadly (at the moment) it doesn't implement NLTH.