So the current scenario is the state of the system (which means we don't care how did we get here). From this we consider each possible move and evaluate all the possible response steps. For the first step this is possible. Now the exchange leads to a new scenario and one check all the memory to see if there is a same one. If so use the previously calculated cost and decide. That cost is coming with distributions since it was the result of a monte carlo search which means play ahead with probability using the same loop (sort of). 累了。