|
July 25th— July 31st |
|
August 1st— August 7th |
|
August 8th— August 14th |
|
This week I learned how we should encode the POMDP examples. But in our current code, the observations depend on the destination state, while in our examples they depend on the source state. Therefore I worked on finding another example with this characteristic which still has a smaller PST learned automaton. I encoded the example and tried on running the code, but I didn’t get any result. When I execute the code, it consumes the whole CPU without giving any output. So I guess it gets into a loop somewhere, and am trying to find where. It is always a big pain to work with other people’s code… |
|
This week we formalized the Merge-Split algorithm, I also worked on labeling the arcs, so that in the learned automaton each arc shows what action/observation pair it corresponds to. And when state s1 gets merged with state s2, all the arcs coming into s1 now point to s2. These changes make the learned machine to look more similar to an automaton, and make the prediction task easier. I also thought about how to add the distinguishablity parameter to the merge condition so that the transitions set do not need to be exactly the same to satisfy the condition. My solution needs an upper bound on the number of states in the original automaton. |
|
Eventually, we found our answer!! This week we found out although Merge-Split learned machine might not be the smallest machine producing the same set of trajectories as the original automaton, it is the minimal predicting machine. Which means when no assumptions are made regarding the start state, the minimum number of states needed to predict if a given trajectory is acceptable in the original automaton or not correctly is equal to the number of states in the Merge-Split learned automaton. Proving this statement made us really happy, because this is in fact what we were looking for. For the next week, I’m going to formally prove the correctness of the Merge-Split algorithm and that it gives the minimal predicting machine. |
|
The proofs are complete and everything is ready to be written in the final paper! I’m going to start writing up my report from this week.
The report will be available in the Final Report section soon.
Here I want to thank CDMP, Doina, Joelle and Prakash for providing me such a wonderful experience during the summer. It was really amazing, I learned a lot from our discussions, and gained very high value skills which I’m certain will be very useful in pursuing my graduate studies.
|
|
| |
|
| |
|
| |
|
Week 1-4 | Week 5-8 | Week 9-12 | Week 13-16 |
|
Dorna Kashef Haghighi |
|
LINKS |
|
August 15th— August 21st |
|
Home | About Me | My Mentor | Project Outline | Journal | Final Report |