The aim of this research effort is to extend the work by Branavan et al. (2012). Our work will introduce the actions of asking questions as part of the agent planning process.
This assumes you have docker
installed on your machine.
$ mkdir tmp && cd tmp
$ wget http://groups.csail.mit.edu/rbg/code/planning/data/env.cache.bz2
$ bzip2 -d env.cache.bz2
$ cd ../
$ sh ./start.sh
$ sh ./start.sh # run the entire system
$ sh ./start_test.sh # run 2 iterations
$ sh ./start_gdb.sh # start with gdb
- Dividing the tasks
- Running Branavan's code (and understanding why it did not stop working after running 1 h)
- Built a suite for testing
-
(issue) There is no space left on the physical machine given by Karthik
-
(success) Discover how to augment the policy with information goals
- (code) We model our questions as PDDL predicates, and load it to the same "possible next subgoal" vector as regular subgoals.
-
(theory) Investigating the policy for predicting the next subgoal
-
(code) After deciding that a subgoal is a question, we need to execute the question, update C before continuing to sample
- (code/theory) For a trivial retrieval system, we can load all the answers in memory
-
(success) 42 million new actions are now down to 517 thanks to Nicola's hardcoding of questions and Adam's hate for thresholds.
-
(success) We found three type of questions:
- Objects (T)
- Subgoal (P*T)
- Comparing two subgoals in the sampled sequence
- Actions (A)
T=50
A=72
P=7
- Finalizing where to add questions and planning how to do so
- Finally getting Branavan's code running
- Setting up the machine given for computation by Regina/Karthik
- Write
Dockerfile
that would compile the code & prepare the environment to run the agent - Setting up
GDB
to simplify C++ debugging
- Understanding the paper and the problem
- Trying to get Branavan's code to work
- Researching on where to add questions
This work is being actively research by Adam Yala, Nicola Greco, and Sebastien Boyer