Michael Buro's 2021-2022 MSc Projects
Mar-8, 2021: I am looking for 3 MSc students for the following
projects - 2021 and 2022 summer funding is secured
Mar-28, 2021: update: added busy beaver project and other project ideas
Project 1
---------
Topic: Learning Motion Tracking via Simulations
Description:
Tracking mobile objects is crucial for effective traffic navigation.
Humans excel at this task and autonomous driving will require it. In
this project we will investigate how to efficiently track objects by
means of supervised learning based on video footage generated by 3d
object trajectory simulators.
Prerequisites:
- Machine learning expertise - in particular deep-network learning
- Familiarity with PyTorch or Tensorflow, Python, and C++
- Computer vision background will also help
=============================================================================
Project 2
---------
Topic: Applying Machine Learning to Variable Selection in SAT Solvers
Description:
Many important real-world combinatorial problems can be reduced to
determining whether a propositional formula has a satisfying variable
assignment. Satisfiability (SAT) solvers have come a long way in
solving large formulas, but haven't used modern machine learning
techniques much.
In this project we will study how deep-neural network learning can be
used to select split-variables more effectively - which has the potential
to speed-up SAT solvers considerably.
Prerequisites:
- Machine learning expertise - in particular deep-network learning
- Familiarity with PyTorch or Tensorflow, Python, and C++
- Theoretical CS background will also help
=============================================================================
Project 3
---------
Topic: Learning pathfinding policies
Description:
In this project we will study how pathfinding algorithms can be
improved by applying deep-network learning to potentially large maps
to train heuristics or motion policies directly from optimal path
training data.
Prerequisites:
- Heuristic Search expertise, in particular pathfinding algorithms
- Machine learning expertise - in particular deep-network learning
- Familiarity with C++ and PyTorch or Tensorflow
=============================================================================
Project 4
---------
Topic: Learning vehicle driving policies
Description:
In this project we will study how local vehicle driving policies can
be learned in a simulated traffic environment, using look-ahead search
and deep reinforcement learning
Prerequisites:
- Heuristic Search expertise, in particular MCTS
- Machine learning expertise - in particular deep-network and reinforcement
learning
- Familiarity with C++ and PyTorch or Tensorflow
=============================================================================
Project 5
---------
Topic: Learning RTS game combat policies
Description:
In this project we will study how large unit battle groups can
cooperate effectively in simulated combat scenarios, based on
look-ahead search and deep reinforcement learning
Prerequisites:
- Heuristic Search expertise, in particular MCTS
- Machine learning expertise - in particular deep-network and
reinforcement learning
- Familiarity with C++ and PyTorch or Tensorflow
=============================================================================
Project 6
---------
Topic: Determining S(5) and ∑(5)
Description:
How long can halting Turing machines with k states and 2 symbols run,
and how many 1-cells can they leave behind on the tape? The related
functions S(k) and ∑(k), defined by Rado in 1962 and also known as
"busy beaver" functions, are non-computable in general, and their
exact values are only known for k ≤ 4. The best known lower bounds for
S(5) and ∑(5) - 4,098 and 47,176,870, respectively - were found in
1990, but the exact values are still unknown.
In this project we will reduce the number of undecided Turing machines
further by developing automated non-termination proofs, with the goal
of reaching 0, and establishing the exact values of S(5) and ∑(5)
Prerequisites:
- Theory: automata, Turing machines, computability, proof techniques
- Programming expertise (C++ or Java preferred)
=============================================================================
Other project ideas to be worked on in collaboration with PhD
students:
A) Speed up best-response computations in imperfect information games
B) Learning to cooperate in multi-player imperfect information games
C) Hierarchical planning in domains with huge state and action spaces