Parallel AI Planning (PAIP)

We will develop the first AI planning system that uses parallel computing to fully exploit modern multi-core processors.

In line with recent hardware developments, today's computers have more processing power and memory than ever before, but individual processors have seen no major speed increases. Yet, artificial intelligence (AI) planning research has mostly stuck to sequential algorithms that fail to take full advantage of these changes. As a consequence, there is currently no commonly-used AI planning system that uses parallel processing. To address this, we plan to create a parallel AI planning system designed from the start for modern multi-core processors. Within it, we will develop efficient parallel algorithms that make the most of modern processors.

Parallelizing existing algorithms on n processor cores usually incurs at most a linear n-fold speedup. This will allow us to solve those tasks faster that are already within reach of sequential planning systems, but it will usually not be enough for solving additional harder tasks. So instead of only computing the same algorithm quicker, we will also develop parallel algorithms that obtain stronger search guidance. This in turn can reduce the size of the explored search space exponentially. Apart from this, the main challenge will be to develop algorithms that scale efficiently to large numbers of processor cores. For this, we need to maintain good thread occupancy by splitting the workload evenly among processors and to minimize the synchronization costs between them.

PI: Jendrik Seipp

Core team: one PhD student, co-supervised by Christoph Kessler

Funding: This project is supported by the Wallenberg AI, Autonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation.