brozek/website
1
0
Fork 0
mirror of https://github.com/Brandon-Rozek/website.git synced 2026-01-30 05:53:38 +00:00
Code Issues Projects Releases Packages Activity

Research page update

Browse source
This commit is contained in:
Brandon Rozek 2026-01-29 15:55:45 -05:00
parent dc51bf1bda
commit 626f0544dd
1 changed files with 31 additions and 15 deletions
Download patch file Download diff file Expand all files Collapse all files

46
content/research/_index.md
View file

@ -7,25 +7,36 @@ Description: A list of my research Projects
**Broad Research Interests:** Automated Reasoning, Automated Planning, Artificial Intelligence, Formal Methods
Currently, I'm a Computer Science PhD Candidate at Rensselaer Polytechnic Institute. I enjoy using logic-based techniques and designing algorithms to solve problems.
Jump to:
- [Planning under uncertainty](#planning-under-uncertainty)
- [Logic](#logic)
- [Symbolic Methods for Cryptography](#symbolic-methods-for-cryptography)
## Planning under Uncertainty
During my PhD I have been primarily focused on investigating planning and sequential decision
making under uncertainty:
- I created a new framework which allows agents to make plans under *qualitative uncertainty*.
This helps in settings where the user doesn't have exact probabilities that various
facts holds, but can instead bucket them into different likelihood values.
This work is supervised under [Selmer Bringsjord](https://homepages.rpi.edu/~brings/).
- Additionally with Selmer Bringsjord in the [RAIR Lab](https://rair.cogsci.rpi.edu/), I have looked at planning through automated reasoning.
I further developed [Spectra](https://github.com/rairlab/spectra) and the underlying
planning with formulas framework to show classes of uncertainty problems that
are easy to encode. Additionally, I wrote a QA algorithm for ShadowProver to integrate to Spectra
for planning under epistemic uncertatinty.
My dissertation topic is on automatically finding and recognizing plans
when agents are uncertain about the environment but can compare the
uncertainty between events *qualitatively*. For example, it is totally
expected when we stack a block that it stays on the top. However, there is a
smaller likelihood that the block falls off.
How can we best make use of this qualitative uncertainty?
- Agents when operating under uncertainty will seek plans which maximize the likelihood of their goals.
I designed an algorithm for recognizing these plans under qualitative possibility theory. This work is supervised under [Selmer Bringsjord](https://kryten.mm.rpi.edu/selmerbringsjord.html) (Paper to be released soon)
- Additionally with Selmer Bringsjord in the [RAIR Lab](https://rair.cogsci.rpi.edu/), I created a framework that captures
situations where agents are able to bucket the likelihood of facts within their environment. I then provide an effective
techinque for using classical planners to find plans which maximize the agent's likelihood of success. ([Paper](/paper/2406.02))
- In the RAIR Lab, I also further developed [Spectra](https://github.com/rairlab/spectra) --
an automated planner built on automated theorem proving. I showed how a class of problems
under uncertainty can be easily encoded and wrote a question-answer algorithm
for ShadowProver so that Spectra can find plans under epistemic uncertainty. ([Paper](/paper/2405.01/))
- With [Junkyu Lee](https://researcher.ibm.com/researcher/view.php?person=ibm-Junkyu.Lee),
[Michael Katz](https://researcher.watson.ibm.com/researcher/view.php?person=ibm-Michael.Katz1),
[Harsha Kokel](https://harshakokel.com/), and [Shirin Sohrabi](https://researcher.watson.ibm.com/researcher/view.php?person=us-ssohrab) at IBM I developed an algorithm
for guiding hiearchical reinforcement agents under partial observability when domain knowledge
can be encoded for characterizing discovery of unknown predicates.
for guiding hierarchical reinforcement agents under partial observability. Specifically,
I focused on situations where the agent knows what they don't know and compiled that knowledge
so that a fully-observable non-deterministic planner can decompose the overall problem. ([Paper](/paper/2406.01))
## Logic
@ -46,7 +57,7 @@ Related Notes:
- [Automated Theorem Proving](atp/)
- [Term Reasoning](termreasoning/)
## Symbolic Methods for Cryptography
Worked with [Andrew Marshall](https://www.marshallandrew.net/) and others in applying term reasoning within computational logic
@ -73,6 +84,11 @@ Collaborators:
Group Website: [https://cryptosolvers.github.io](https://cryptosolvers.github.io)
---
**Note:** From this point on, the projects listed happened over 5 years ago.
---
## Reinforcement Learning