pip3 install -e gym-pursuit-evasion
By default this runs in the random evader environment. Running the command below, has an inference based agent catch an evader in a simple grid based environment. The agent is incentivised to do this quickly due to the reward for each time step when the goal is not reached to be -10 and the reward for reaching the goal = 0. Note for the inference based agent, rewards must be negative due to the paradigm we are using which is explained here: https://arxiv.org/pdf/1805.00909.pdf
python3 -W ignore agent-dice.py
By default this runs in the stationary evader environment. This command executes an agent that is modeled identical to the previous command but is written in Pyro
python3 -W ignore agent-dice.py
The py2dice directory contains the code for dynamically creating ASTs in Dice using Python and using them to progammatically generate Dice programs using Python. The Dice Program Generator in dice_inference_engine.py however for simplicity doesn't rely on this, but this tool is more generally applicable and will hopefully serve as a way to extend the functionality of Dice.
- The tmp directory is where the videos generated from executing the agent commands are stored. This can be useful to observe how the agent is doing.
- The .infer directory is where the dynamically created Dice programs are stored.
- The gym-pursuit-evasion directory contains the Open AI Gym Wrapped Environment created for Reinforcement Learning in this project.
- The dice_scratch_code directory contains the scratch code written in designing the Dice Inference Engine
- The pyro_scratch_code directory contains the scratch code written in designing the Pyro-based agent