5.3 Primitive Actions

Figure 9 shows that the improved runtime performance of the community is a direct result of the fact that the planner produces plans which are either more efficient (e.g., interleaves goals) or more likely to be successful or both. The left side of the figure tracks the number of actions attempted (whether successful or not) by the actors over the course of solving a problem. The right side focuses on just the number of successful actions. Learning conventions leads to significantly fewer actions, both attempted and successful. Standard deviations for all points shown are less than 0.82 and, again, there is no significant decrease after the third problem-solving episode.

By focusing on the performance of the baseline system, we can see that, over time, the operator probability trees lead the individual actor to produce plans that require fewer attempted actions. However, this is not translated into a reduction in the number of successful actions needed to solve the problems. So the reduction in the number of rounds for the baseline system reflects the fact that actors find the appropriate kinds of first-principles plans earlier in the course of the activity. On the other hand, procedural memory allows the actors to solve the problems using fewer successful actions. Clearly, the actors are guided by higher caliber plans.

  

Figure 9: The number of primitive actions is decreasing.

Figure 9 takes a macroscopic view, measuring the action effort of the community to solve the entire set of top-level goals. Alternatively, we can myopically measure how efficiently the actors are solving individual top-level goals in isolation. For a given goal, we can compute the number of actions, either attempted or successful, that the actors undertook to achieve the goal. This can be compared to the fewest possible given the initial configuration of the problem. For example, a small, clear box inside the house requires a minimum of 3 actions to get it onto the truck: LIFT, CARRY and LOAD. Then, for all problems and all top-level goals, the frequency with which the community undertakes the minimal number of actions can be computed. The results are given in Figure 10. With procedural memory, the actors in MOVERS-WORLD are attempting the minimal number of actions more frequently (80.6% of the time) than in the baseline system system (60.3% of the time). Also, despite a preliminary dip below baseline performance, the actors solved problems using the minimal number of actions more frequently when they learned conventions.

  

Figure 10: Approaching minimal action effort.