Diadem is a python library for development and training of interaction-aware decision making algorithms based on tensorflow<2.0. The focus is on combinations of Deep Reinforcement Learning with classical AI techniques, such as MCTS. Currently, the training and evaluation framework and DQN variants (prioritized experience replay, Learning from Demonstrations (DQfD), C51, Quantile Regression DQN) are implemented. With the environment wrapper class, the library allows integration of different environments, standard benchmarks such as OpenAI Gym but also bark, focusing on interactive decision making in multi-agent scenarios.

• Bazel (requires Java)
• Python3
• virtualenv

1. Clone the repository and change to base repository directory
2. bash scripts/setup_venv.sh: this will create a virtual python environment (located in ./venv) and install all necessary python dependencies.
3. source scripts/into_venv.sh: this will activate the virtual environment (keep this in mind for the future: each time you use Bazel, even beyond this installation, be sure to have run this command beforehand)
4. bazel test //...: this will run all specified tests. All tests should pass successfully. (individual tests can be executed e.g. with bazel test //examples:dqn_basic
5. For online rendering of the environment use the run mode, e.g for the GymCartpole environment and DQN use bazel run //examples:dqn_basic

The examples folder shows how to run the DQN variants with the OpenAIGym environment wrapper. The configuration files extend from the common configuration by importing the base property from the common parameters file. This allows for a simplistic parameter definition for specific use cases based on a common parameter set avoiding parameter repetitions.