Betreuer/in: Baron
Time-Sensitive Networking (TSN) extends the IEEE 802.1 Ethernet standard with
mechanisms for deterministic, low-latency communication, making it a cornerstone
technology for industrial automation, autonomous driving, and safety-critical
applications [1], [2].
Schedule synthesis, the assignment of precise transmission
windows to time-triggered streams across a network, is a computationally hard problem
[3] that makes exact optimization methods impractical for large deployments. The
ITANS framework, developed by Arestova et al. [4], addresses this challenge using a
genetic algorithm (GA) that searches the space of valid stream orderings to minimize
network makespan. ITANS applies a fixed operator configuration, Linear Order
Crossover (LOX) and Swap mutation, throughout every generation of every run,
regardless of how the search is progressing.
We present RL-ITANS, an extension of ITANS in which a tabular Q-learning agent [5], [6]
observes the state of the GA after every generation and dynamically selects which
operator configuration to apply next, choosing among five configurations that vary in
crossover type, mutation type, and mutation rate. The agent learns a state-dependent
selection policy across many successive GA runs, accumulating experience in a
persistent Q-table. The implementation extends ITANS’s Java codebase minimally,
without modifying any of the original source files, and communicates with a Python-
hosted learning server via a lightweight HTTP protocol.
An experimental evaluation compares RL-ITANS against the static-operator baseline
using 3,494 one-to-one matched pairs of runs, each pair sharing an identical underlying
topology instance, drawn from 24 topology configurations spanning stream counts of 10
to 800 and bridge counts of 5 to 30. Results show no practically meaningful difference
in schedule quality (makespan): differences at every stream count are well under a
tenth of a percent and orders of magnitude are smaller than the within-method
standard deviation. Wall-clock duration results depend strongly on topology size: for
topologies of 100 streams or more, the scale at which per generation computation
dominates communication overhead, RL-ITANS completes faster than its matched
baseline in 70 to 83 % of pairs, with a mean per-pair time saving of 11 to 21 %; at the
smallest topologies (10 to 20 streams) this advantage reverses sharply, with RL-ITANS
running 47 to 97 % slower on average.
Room 04.137, Martensstr. 3, Erlangen
or
Zoom:
https://fau.zoom-x.de/j/68350702053?pwd=UkF3aXY0QUdjeSsyR0tyRWtLQ0hYUT09
Meeting-ID: 683 5070 2053
Kenncode: 647333