Section 5. Learning to Interleave Goals

As mentioned previously, using execution traces as a basis for future cooperative plans allows the agents to learn plans beyond the scope of the first-principles planner (or a traditional second-order planner based on it). Another example shows how lifter L2 learns to interleave two goals. In this example, HTO and L1 will act as in the first example, with the box in question now extra large box XLBOX1 rather than MBOX3. The following are L2's plans and actions, starting four steps before HTO makes her first request to L1.

1.

L2 creates a typical plan to get XLBOX1 onto the truck: to lift, carry and load the box jointly with L1. L1 agrees to lift the box together and they do so. L1 then agrees to carry the box to the street. However, XLBOX1 is too large to carry, even jointly, and the operator (and hence rest of the plan) fails.

2.

L2 creates a plan to get SBOX5 onto the truck. The plan consists of putting down XLBOX1 with L1's help and then lifting, carrying and loading SBOX5 onto the truck by herself. L2 is delayed in asking for L1's assistance because HTO calls first with a request to put XLBOX1 onto HANDTR2. L1 agrees to help HTO.

3.

When L2 does ask L1 to help achieve HANDEMPTY via putting XLBOX1 down together, L1 replies that she would rather load the box together onto the hand-truck. L2's planner adapts her current plan by replacing the PUT-DOWN-TOGETHER with the appropriate LOAD-TOGETHER. The agents then load XLBOX1 onto the hand-truck. L2 continues on with her plan and loads SBOX5 onto the truck. Meanwhile, HTO has pushed the hand-truck to the street and L1 has agreed to get XLBOX1 onto the truck.

4.

L2 constructs a plan to get large box LBOX2 onto the truck and moves back to ROOM1.

5.

L2 is interrupted before attempting to lift LBOX2 by a request from L1 to help unload XLBOX1 from the hand-truck. L2 constructs the plan of moving to the street and then unloading XLBOX1. The agents do so.

6.

L1 asks L2 to load XLBOX1 onto the truck and they do.

The sequence of actions L2 undertakes corresponds to six different calls to the planner. Nonetheless, L2 can extract a single cooperative procedure from her execution traces by the machinery of collective memory (showing the original literals instead of new variables for clarity):

((WAIT-FOR L1 (ON XLBOX1 HANDTR2))
 (LOAD-TOGETHER XLBOX1 HANDTR2)
 (LOAD SBOX5 TRUCK1)
 (WAIT-FOR L1 (ON XLBOX1 TRUCK1))
 (LOAD-TOGETHER XLBOX1 TRUCK1))

Interleaving loading and unloading a box onto the hand-truck with loading a small box onto the truck, as in this example, is the most useful procedure lifters learn in the current version of the system. Its utility arises from low idle time and wide applicability.