Fast Downward Cedalion

Cedalion is our algorithm for automatically configuring sequential planning portfolios. Given a parametrized planner and a set of training instances, it iteratively selects the pair of planner configuration and time slice that improves the current portfolio the most per time spent. At the end of each iteration all instances for which the current portfolio finds the best solution are removed from the training set. The algorithm stops when the the total runtime of the added configurations reaches the portfolio time limit (usually 30 minutes) or if the training set becomes empty. Conceptually, Cedalion is similar to Hydra (Xu, Hoos, and Leyton-Brown 2010) in that both use an algorithm configuration procedure (Hutter 2009) to add the most improving configuration to the existing portfolio in each iteration. However, Hydra uses the algorithm selection system SATzilla (Xu et al. 2008) to select a configuration based on the characteristics of a given test instance, and therefore does not have a notion of time slices. In contrast, Cedalion runs all found configurations sequentially regardless of the instance at hand and makes the length of the time slices part of the configuration space. Cedalion is also very similar to the greedy algorithm presented by Streeter, Golovin, and Smith (2007). Given a finite set of solvers and their runtimes on the training set, that algorithm iteratively adds the (solver, time slice) pair that most improves the current portfolio per time spent. In contrast, Cedalion does not rely on a priori runtime data and supports infinite sets of solver configurations by using an algorithm configuration procedure to adaptively gather this data for promising configurations only. In principle, Cedalion could employ any algorithm configuration procedure to select the next (configuration, time slice) pair. Here, we use the model-based configurator SMAC (Hutter, Hoos, and Leyton-Brown 2011) for this task. As a simple standard parallelization method (Hutter, Hoos, and Leyton-Brown 2012), we performed 5 SMAC runs in parallel in every iteration of Cedalion and used the best of the 5 identified (configuration, time slice) pairs. We could have included planners other than Fast Downward in our Cedalion portfolios (even other parameterized planning frameworks, by configuring on the union of all parameter spaces). If portfolios prove to be useful in the learning track setting, this would have almost certainly improved performance, due to the fact that portfolios can exploit the complementary strengths of diverse approaches. Nevertheless, we chose to limit ourselves to Fast Downward in order to quantify the performance gain possible within this framework.