pachinko is a package that implements one-shot optimal decision-making agents to solve environments provided by the truman package. The goal is to implement agents that learn in a single experience, so that they can perform well on unique systems that can't be reliably simulated and that have a high cost of experimentation.
For more information about the types of environment provided by truman, see the truman package README.
pachinko implements various algorithms exposed as agents compatible with OpenAI Gym environments:
- the Random agent, which takes a random action each timestep
- the EpsilonGreedy agent, which takes the highest conversion rate action, with a small probability of choosing a random action
- the Periodic agent, which splits the time series into a repeating period (eg a separate conversion rate for each day of the week) and chooses an action each step based on the upper confidence bound of the conversion rate
pachinko is light on testing - the true test of the agents is their performance in environment suites from the truman package.
To get started, you'll need to have Python 3.7+ installed. Then:
pip install pachinko
You can also clone the pachinko Git repository directly. This is useful when you're working on adding new agents or modifying pachinko itself. Clone and install in editable mode using:
git clone https://github.com/datavaluepeople/pachinko
cd pachinko
pip install -e .