Experiments for PFTDPW & SparsePFT

Generalized Optimality Guarantees for Solving Continuous Observation POMDPs through Particle Belief MDP Approximation

Reproducing experimental results

Each experiment contains a compare.jl script that takes iter and addprocs flags.

• iter dictates how many monte carlo evaluation runs to perform for each planning time and solver.
• addprocs dictates how many additional workers to add.

For example, if we want to reproduce the experimental results for Laser Tag, from the experiments folder we could run the following:

julia --project LaserTag/compare.jl --iter=500 --addprocs=30

.

The results for experiments are stored in experiments/<experiment_name>/data/.

Hyperparameter optimization

Hyperparameter optimization scripts are stored in experiments/<experiment_name>/opt/.

Example hyperparameter optimization results for PFT-DPW on Sub Hunt: