A library for composable explainability in reinforcement learning agents.

Reading explanations of other people's RL agents and giving explanations of our own sucks. We want to write and train our RL agents and then deploy that into a front-end for others to explore.

No more crawling through Python code and more focus on seeing algorithms in action.

Zalia Flow (zaliaflow.io) is an enterprise toolkit for building artificial intelligence that people delight in trusting. Our core libraries are released as open-source software.

Right now, you can install from source and run an example.

$ git clone git@github.com:ZaliaFlow/decode-py.git
$ cd decode
$ python3 -m pip install -r requirements.txt
$ cd src
$ streamlit run example/pong.py

In example/pong.py is the hello-world of reinforcement learning: a random agent sampling an environment.

Right now, there are two lines that matter:

memory = Memory(engine = Memory.make.new(capacity = 1000).unwrap())
...
memory.update(parent = memory_display, children = None).unwrap()

This is a high-level API for getting an experience buffer for Deep Q-Networks that will open up a display of the agent's results from sampling the reward distribution.

Help us find out what's the single most impactful view of your machine-learning code that would make it effortless to explain.

Book a meeting with our team on https://www.zaliaflow.io/book-online to chat.