Hierarchical Organization of Robots

6 Results

In the simplest cases, where the digit strings are to base 2, all strings can be observed and there is at least one action affecting each digit string and the probabilities of multiple effects of actions and interaction among state digits is low, then no hierarchical structure is required and every deviation of a state string digit from its target is corrected within one period.

Increasing the number base but leaving the other parameters of the simulation model unchanged creates the need for planning but again there is no gain from hierarchical organization The number of time periods needed to eliminate deviations of observed digit values from their target values is not changed by introducing hierarchical relations.

We already know from experiments in computational organization theory that, when there is substantial incidence of multiple effects of individual actions, the chief executive of the group can reorganize observing agents to increase efficiency. What happens is that actions are coordinated for their effects.

The hypothesis to be tested is whether the effects of cumulative malfunctioning of robots (i.e. there are no repair services) can be offset by periodic reorganization. We investigate the effects of organizing agents each with simple mental modelling capacities to determine whether cascades of simple models are able to capture the relevant relationships which exceed the capacity of one agent to capture. The approach to be taken follows from Ye and Carley (1995) who set up a model in which superiors guess the true state of the environment based on the judgements of their subordinates. They are in effect modelling subordinates.

The demonstrator model yielded the pattern of errors over 10 dates (40 task cycles) where there was a probability of 0.2 that a 0 would change to a 1 is given in fig. 1.

fig. 1: Time pattern over elapsed task cycles of the number of 1s in the state string with no reorganization of tasks. From the 21st task cycle onwards, either 35 or 36 digits took the value 1 out of a possible 36.

When reorganization is allowed the observed pattern of errors is as given in fig. 2.

fig. 2: Time pattern over elapsed task cycles of the number of 1s in the state string with reorganization of tasks within rooms every 4th task cycle.From the 22nd task cycle onwards, averaging about 30 digits taking the value 1 out f a possible 36 from the 32nd task cycle onwards.

Clearly, the results shown indicate that in these experiments, the rate at which individual robot malfunctions affect overall robot-team performance is significantly reduced by reorganization of tasks.

Hierarchical Organization of Robots - 20 MAY 98

[Next] [Previous] [Top] [Contents]

Generated with CERN WebMaker