rl-psr

Code for "Reconciling Rewards with Predictive State Representations", A. Baisero and C. Amato, IJCAI 2021.

The following packages are required, and can be found on github:

• rl-parsers <https://github.com/abaisero/rl-parsers>_
• gym-pomdps <https://github.com/abaisero/gym-pomdps>_

Preferably, you should be creating an environment specifically to install the required packages first, then this one, and then run the experiments. For each of the above, in that same order:

1. move to the package directory, and install using

   python -m pip install .
   

2. run the tests to make sure installation was correct

   python -m unittest discover
   

The code to replicate the experiments from the IJCAI 2021 paper is contained in the experiments/ folder. Move there, then run the following commands. NOTE: in these scripts, bsr stands for belief-state representation (i.e. POMDP), psr stands for predictive-state representation, and rpsr stands for reward-predictvie state representation.

File pomdps.all.txt contains all the POMDPs tested to verify the significance of the accuracy problem of PSRs; while pomdps.ijcai21.txt contains the 6 domains whose PSRs are non-accurate, for which we could run value iteration, and which form the basis for the more thorough evaluation.

1. Search for core sets of PSRs and R-PSRs:

   <pomdps.ijcai21.txt ./search.local
   

   This will compute core sets, print their ranks, and store them in cores/.

2. Compute reward errors of PSRs and R-PSRs w.r.t. the POMDP rewards:

   <pomdps.ijcai21.txt ./info.local
   

   This will compute error measures, print them to standard output, and store them in infos/.

3. Run POMDP-VI, PSR-VI and R-PSR-VI:

   <pomdps.ijcai21.txt ./vi.local
   

   This will run the value iteration algorithms for 150 iterations, and store the resulting value functions in vfs/. This is the slowest step; it will take many hours if a single machine is used.

4. Plot a quasi-Bellman-error measure to check for convergence of the value functions:

   <pomdps.ijcai21.txt ./plot.local
   

   This will plot the convergence properties of the value functions, and store them in plots/.

5. Evaluate the value functions' respective policies:

   <pomdps.ijcai21.txt ./eval.local
   

   This will run the Random, POMDP-VI, PSR-VI and R-PSR-VI policies for 100 episodes of 1000 steps each, calculate the true and estimated returns, and store them in evals/.

6. Compile the evaluation results into tables:

   <pomdps.ijcai21.txt ./tables | tee tables.tex
   

   This will aggregate the results obtained by the evaluation step, print the results in a tex/table format, and save the results in tables.tex.