- Download the repository and unzip the planning domain file "spudd_sperseus.zip" under the current directory.
- Install the required packages
pip install -r requirements, supportpython 3.8and above. Note that by default this will install jax that supports cuda. However, this is not required but only a speed up for the backward loopy belief propagation algorithm. - To run the experiments run
python3 main.py -i <instance full name> -a <algorithms to run>. For example,python3 main.py -i sysadmin_inst_mdp__1.spudd elevators_inst_mdp__5.spudd sysadmin_inst_mdp__2.spudd -a mfvi_bwd fwd_lbp random. If given no parameter, the program will run all 60 instances from 6 domains on all algorithms (mfvi_bwd, mfvi_fwd, mfvi_bwd_noS, bwd_lbp, fwd_lbp, csvi_bwd, csvi_fwd, exp_mfvi_bwd, exp_csvi_bwd).- For running the demo problem, use (demo_mfvi_bwd, demo_mfvi_bwd_med, demo_mfvi_bwd_noS, demo_mfvi_bwd_main_all, demo_csvi_bwd, random) for algorithm parameters to get the detailed plotting and print out. The demo instance names include (demo_inst_mdp__1, demo_inst_mdp__2, demo_inst_mdp__3). To regenerate the ELBO plotting in other problem domains, please run the corresponding demo version of the algorithm on the problem instance.
- To visulize the results
python3 summary.py -a <algorithm to visualize>will generate the summary comparison plots among all the domains, to facilitate plotting, please put 'random' as the last algorithm.python3 visual.py -d <domain> -a <algorithm to visualize>will generate the algorithm comparison plots of the domain. If given no parameter, it will generate plots on all 6 domains of algorithms mfvi_bwd, csvi_bwd, fwd_lbp, sogbofa, random.
- Please refer to
python3 main.py -h,python3 summary.py -h,python3 visual.py -hfor more information. - All the algorithmic and expermental parameters are stored in
./conf/alg_param.jsonand should be modified to align with running time and computational resourse contrains. - The results in the paper are stored in
./results_of_paper.zip. To visulize based on those results, please unzip it into the folderresults. - To compare with sogbofa, please unzip
sogbofa.zipand put the results of sogbofa inside./results/. - To regenerate the results of sogbofa, see the README file inside the SOGBOFA directory.
- Notice that the code does not erase the previous running results in order to faciliate parallel running and data collection. Thus running the same algorithm on the same instance multiple times will accumulate the results in the same file, which might cause problems during visualization if the algorithms under comparison have a different number of results. To avoid it, please clean the
resultsfolder every time before running. - To visualize CSVI w.r.t iteration number, one convenient way is to use the original results in
results_of_paper.zip. You could also generate new results by tuning parameters in./conf/alg_param.json, but need to put the results in./results/iter_comp/iter1and./results/iter_comp/iter100respectively to make the visulization work. - Before running, for things to work normally, the directory structure should look like (The
logsdirectory is optional and will be automatically generated and theresultsdirectory is optional if not using the pre-generated results.) - If you run into memory issue in running backward loopy Belief Propagation, that is because jax by default preallocates the memory. Please refer to Jax Doc for corresponding modification or switch to CPU instead by setting
jax.config.update('jax_platform_name', 'cpu')inutils.py.
project
|
│ README.md
│ requirements.txt
| results_of_paper.zip
| sogbofa.zip
| spudd_sperseus.zip
|
└───conf
│ │ default.conf
│ │ alg_param.json
| | ...
└───results
│ └───alg1
│ │ <instance>_<alg1>_rlt.json
│ │ ...
│ └───alg2
| │ ...
| ...
└───spudd_sperseus
│ │ <instance>.spudd
│ │ ...
└───src
│ │ main.py
│ │ ...
└───logs
│ └───alg1
| | | ...
- The code of second forward loopy Belief Propagation is based on and modified from Aleksei Krasikov's github repository.
- The code of backward loopy Belief Propagation is based on PGMax.
- The code of SOGBOFA is based on and modified from Hao Cui's github repository.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent