rejuvyesh / SMM_experiments

Reproducing Experiments with Structured Mechanical Models

Glossary

Naive refers to BBNN in the paper. ControlAffine refers to SMM-C in the paper.

Pre-requisites

1. Install GNU Parallel.

On Ubuntu it should be as easy as:

sudo apt install parallel

2. Setup julia project

julia --project --color=yes -e 'using Pkg; Pkg.instantiate(); Pkg.build(); Pkg.precompile()'

3. Dump validation data

parallel -j4 ./dumpvaliddata.jl ::: furuta cartpole acrobot doublecartpole ::: "runs/"

Training

Furuta Pendulum

parallel -j5 ./train.jl \
    ::: "ControlAffine" "Naive" \
    ::: furuta \
    ::: 8192 4096 2048 1024 512 256 \
    ::: 42 1234 1339 178969 1234321 \
    ::: 2000 \
    ::: 256 \
    ::: "runs/"

parallel -j5 ./train.jl \
    ::: "Naive" \
    ::: furuta \
    ::: 16384 32768 \
    ::: 42 1234 1339 178969 1234321 \
    ::: 2000 \
    ::: 256 \
    ::: "runs/"

Cartpole

parallel -j5 ./train.jl \
    ::: "ControlAffine" "Naive" \
    ::: cartpole \
    ::: 8192 4096 2048 1024 512 256 \
    ::: 42 1234 1339 178969 1234321 \
    ::: 2000 \
    ::: 256 \
    ::: "runs/"

parallel -j5 ./train.jl \
    ::: "Naive" \
    ::: cartpole \
    ::: 16384 32768 \
    ::: 42 1234 1339 178969 1234321 \
    ::: 2000 \
    ::: 256 \
    ::: "runs/"

Acrobot

parallel -j5 ./train.jl \
    ::: "ControlAffine" "Naive" \
    ::: acrobot \
    ::: 8192 4096 2048 1024 512 256 \
    ::: 42 1234 1339 178969 1234321 \
    ::: 2000 \
    ::: 256 \
    ::: "runs/"

parallel -j5 ./train.jl \
    ::: "Naive" \
    ::: acrobot \
    ::: 16384 32768 \
    ::: 42 1234 1339 178969 1234321 \
    ::: 2000 \
    ::: 256 \
    ::: "runs/"

Double Cartpole

parallel -j5 ./train.jl \
    ::: "ControlAffine" \
    ::: doublecartpole \
    ::: 65536 16384 8192 4096 2048 \
    ::: 42 1234 1339 178969 1234321 \
    ::: 2000 \
    ::: 256 \
    ::: "runs/"

parallel -j5 ./train.jl \
    ::: "Naive" \
    ::: doublecartpole \
    ::: 262144 131072 65536 32768 16384 8192 4096 2048 \
    ::: 42 1234 1339 178969 1234321 \
    ::: 2000 \
    ::: 256 \
    ::: "runs/"

Evaluation

Run Evaluation

Get best Validation Error from logs

./scripts/get_best_validation_error.py <path to logs> > val_error.json

Get the control costs

./evaluate.jl <Naive or ControlAffine> <taskname> cost <path to `best.th` generated from training>

Collect results

./scripts/collect_val_error.py <path containing dumped val_error.json>

./scripts/collect_control_costs.py <path containing dumped cost.json>