SPLIT-TRACE

The purpose of this function is to identify chunks of operators; each chunk of operators must collectively accomplish a clearly-defined set of top-level goals. It currently identifies two kinds of chunks, but in the future, additional techniques may identify other types of chunks. SPLIT-TRACE takes a summarized execution trace as input and returns a list of operator-chunk/goals pairs as output.

Using the active goal information recorded in the trace events, SPLIT-TRACE considers a consecutive subsequence of the (chronologically-sorted) summary to be a chunk if the joint effects of the executed operators achieves the conjunction of the goals associated with them. SPLIT-TRACE combines these ``temporal'' chunks with ``goal-directed'' ones, which are determined by simply grouping together operators associated with identical goals.

  

Figure 6: Schematic of a temporal chunk.

These two types of chunks are represented schematically in Figure 6. The axis represents a portion of an agent's activity timeline with markers indicating the execution of actions A1 through A9 (more actions may precede or follow these actions). If the agent, individually or in cooperation with other agents, accomplished goals G1 and G2 during the intervals marked on the timeline, there would be three chunks identified from this activity timeline. There would be a goal-directed chunk for G1 of (A2, A3, A7, A8), a goal-directed chunk for G2 of (A4, A5, A6) and a temporal chunk for both goals of (A2, A3, A4, A5, A6, A7, A8). This last chunk is the simplest type of chunk which is a temporal chunk that is not also a goal-directed chunk; a similar chunk forms the basis for agents to learn to interleave goals in the example in Section 5.