Task-Guided Inverse Reinforcement Learning Under Partial Information

​

Dependencies

​ This manual has been tested on a clean Ubuntu 20.04 LTS installation. ​

• The packages has been tested on Python 3.8 with numpy, gurobipy, stormpy installed.
• For visualization matplotlib is needed, in addition to tikzplot ​

Install the package

To install the package, go to the directory MCE_IRL_POMDPs where it has been extracted and execute

python3 -m pip install -e .

​

Reproducibility Instructions

​ We provide the command in order to reproduce the reulst in the paper. Note that the directory

examples/all_domains/

contains all the POMDP model descriptions you want to reproduce the expected reward and computation time of SolvePOMDP and SARSOP ​

Table 1.: Comparison with existing approaches

In order to obtain the computation time and reward in all the domains by our approach you need to execute the file bench_scpforward.py

python3 bench_scpforward.py

​ Comment and uncomment lines in this file according to the benchmark instance you want to reproduce, ​

Influence of side information

​ Execute the following commands to learn a policy and plot the policy. The following command can be executed for every other envronemnts such as avoid_example, evade_example ​

python3 final_exp_maze.py
python3 plot_maze_result.py