forainactionsdo s0,r env.STEP(a) solution.APPEND(a) s s0 ifsolution.LENGTH()>Lathen returnNone ifenv.SOLVED()then returnsolution returnNone functionPLANNER(state)

Therefore,toensureasimilar computational budget, we limit the number of planner calls toLp = 8 for MCTS-kSubS and to Lp = 24forthe baseline -sothe number ofstates visited overthe course ofasingle solverrun is similarforbothmethods. Top-left part of Figure 1 illustrates results of MCTS experiments. For every number of planning passesP, MCTS-kSubS has significantly higher success rate than the corresponding baseline experiment. To speed up training and inference we use its lightweight version. Preparing data points for the training of the generator is described in Algorithm 8.