HierarchyCraft Environment Benchmark

Benchmarking of classic flat reinforcment learning agent (FRL) and hierarchical reinforcement learning (HRL) agents on the HierarchyCraft environment.

Research objective

This work aims at evaluating the number of interaction steps and learning steps needed to succeed at hierarchical tasks provided by the HierarchyCraft environement (performance) in order to evaluate the following hypotheses:

1. HRL agents performs that FRL on some hierarchical tasks.
2. Agents performance on a task is correlated to the either the learning or total complexity of given hierarchical behavior explanations (HEB) of the task as defined in the HEBGraphs framework.
3. The performance of an agent on a new task can be predicted using the correlation between performance and total or learning complexity on previous tasks.
4. The difference of performance between HRL and FRL agents on a new task can be predicted using the difference of correlation between total and learning complexity for each agent on previous tasks.

Installation

Install from GitHub using pip

pip install git+https://github.com/MathisFederico/HierarchyCraftBenchmark.git

Run the benchmark

craftbench