markus semmler
i work on things that have to make good decisions while not being sure what’s going on.
ml, reinforcement learning and probabilistic inference engineer.
available for remote work with overlap for eu and us hours.
p(a, o | s) = π(a | s) p(o | s, a)
work
rl & control
| project | what | link |
|---|---|---|
| policy gradient methods | trpo and mpo in pytorch, tested on pendulum and quanser sims | code |
| device calibration | rl agent, sim to real | |
| set-point control | model predictive control for an industrial machine |
probabilistic
| project | what | link |
|---|---|---|
| multi-object tracker | bayesian, kalman filter, loopy belief propagation | |
| forecasting and pricing | bayesian filter, demand forecasting |
data & llm
| project | what | link |
|---|---|---|
| influence functions | batched, for neural nets | |
| log anomaly detection | llm for anomaly detection in logs | |
| classwise-shapley | tmlr reproduction, certified | code |
stack
inference: state-space models, bayesian filtering, kalman filter, variational inference, mcmc, factor graphs, belief propagation
rl & control: policy gradient methods, sim-to-real, model predictive control / ilqr, active inference
tools: jax, pytorch, ray, c++, python
papers
- fisher information approximations in policy gradients
- [re] classwise-shapley values for data valuation, tmlr, reproducibility certified · code
contact
open for contract work. marsem.dev [at] gmail [dot] com
· github
· freelancermap
m.sc. computer science, tu darmstadt. posterior updated daily.