wait for (done,t,outCount,b,r) from topmost coord. tnext := t busy := b rollback := r

If a causality error has been detected the root coordinator initiates the execution of the rollback through a rollback-message. Otherwise the simulation proceeds as usual. There is another problem the root coordinator has to handle: if there are deliberation processes running and no events are scheduled, i.e. tnext = has been returned to the root coordinator. In this case, the root coordinator sends a *-message with an estimated tnext , typically a number close to infinity. This will cause the simulator to wait until at least one of the deliberation processes has finished and another event (completion of the process) can be scheduled.

1

5 Evaluation: Agents in T ILEWORLD
Initially, T ILEWORLD was developed to test different control, particularly commitment, strategies of IRMA agents [9, 8].
1 1 2 3 4 5 6 7
B (1) 6/2 A (2) 4/3 B (3) 10/5

2

3

4
A (3) 9/6

5
TGA

6

7 A

C A B A A A A
C (2) 10/8

OGA

C A

A A B

B
C (3) 8/6

B (1) 10/7

A

search space and implies a costly deliberation with respect to computing time and memory. The T ILEWORLD scenario we have chosen comprises an 8 by 8 grid, 1000 units of simulation time, and a real-time knob, i.e. factor, of 1. Thus, 1 unit of simulation time should be about 1 second. The grid elements change, e.g. holes and tiles appear and disappear, every 50 time units with a probability of 40%. All agents we tested had a scan range of 5 grid elements, limited, but sufficient fuel, and were planning for two goals simultaneously. Within our implementation of T ILEWORLD scanning requires intensive message exchange. The experiments were run on 2 Ultra 2 machines equipped with about 200 MB each. Each experiment consisted of 15 runs. We first put our algorithm to test using one single agent in the T ILEWORLD. The time the simulation runs needed to complete averaged slightly less than 1200 seconds. About 150 of those were due to the scanning activity, and more than 900 were due to planning. Afterwards, we added another agent to the scenario. For the experiment JAMES distributed the model and the processor tree. The agents and their simulators, including their planners, were running on different machines. Each of the agents was planning an average of more than 900 seconds. The total time used for the simulation averaged about 1450 seconds. About 250 of those were due to the scanning activity of the agents. Thus, the 900 additional seconds of planning time for the second agent caused almost no additional overhead in simulation time. The overhead caused by the sending of rollback messages turned out to be negligible. However, we did not measure the effort required for saving the state of the model. Not surprisingly, running two agents on a single machine, requires about twice the computation time. The scanning is of course faster but this does not compensate for the loss of efficiency caused by the sequential execution of the planners.

Figure 3. A T ILEWORLD scenario [10] T ILEWORLD (Figure 3) is a two dimensional grid world with tiles, which can be moved, and holes, which should be filled with tiles. There are obstacles, which impede the movement of agents, and gas stations which allow the refilling of consumed energy. Tiles, holes, and obstacles appear and disappear at certain rates, according to global parameter settings. Thus, the environment displays probabilistic, dynamic behavior. The effectiveness of an agent is measured in terms of scores that summarize the number and kind of holes filled, and the type of tiles used for filling. T ILEWORLD combines a counting problem, how many more tiles of what type does the agent need to fill a particular hole, with route planning in a grid world. This setting puts only few constraints on the

6 Conclusion
The testing of multiple, deliberative agents is space- and time consuming. External modules are plugged into a frame provided by the test bed. The frame provides the interface between agent and agent-architecture to be tested and the virtual environment agents shall be tested in. Since the performance of agents depends significantly on their timely decisions, a time model is employed to relate the actual or expected execution time of agents to the virtual time of the test environment. One time model is often applied due to its flexibility and simplicity: it clocks the execution of the deliberation component and applies a function to transform the consumed time into simulation time. Thus, only after the generation of a plan the simulation will