Up: EVALUATING LONG-TERM DEPENDENCY BENCHMARK
Previous: FINAL REMARKS
Previously proposed long time lag algorithms can do significantly
worse than simple RG for certain
already proposed benchmark tasks. In other cases (e.g.,
involving subtasks such as pattern recognition in continuous
spaces, or additional input noise), RG does not do as well.
This suggests that simple-minded strategies such
as RG should be routinely included among algorithms against
which newly introduced long time lag algorithms are compared.
Furthermore, new benchmark tasks should be designed to make
simple random search algorithms fail on them.