Planning through Backpropagation

This is a refined version of Tensorflow planner on planning problem.

In the training stage, we train the transition functions through the previous observations. In another words, we assume the trainsition function is unknown while the reward function is given.

The code is able to connect to the RDDL simulator by calling python through commandline tools.

Example

Train

python train.py \
-p data/res/reservoir4/ \
-x Reservoir_Data.txt \
-y Reservoir_Label.txt \
-w weights/reservoir/reservoir4 \
-s 4 \
-d Reservoir

Plan

python plan.py \
-w weights/reservoir/reservoir3 \
-d Reservoir \
-i Reservoir3 \
-s 3 \
-a 3 \
--initial temp/state

More concrete examples could be found in Commands.md file.

Note:

1. the initial state is optional, and the default is zero state.
2. Action constrain need to manually set before running planner.