It is a collaborative project with Prof. VS Subrahmanian and Prof. Sarit Kraus. We study that when a set of agents are deployed over a dynamic network, a major problem is the deployed
multiagent applications may easily crash due to external events. The ability of multiagent systems to provide services in the presence of attacks and failures is known as the survivability of multiagent systems.
Replicating agents on different hosting nodes is a natural and efficient way to ensure the survivability of multiagent
systems.
In most of existing replication approaches, it is common that
there is one special algorithm in the network which is in
charge of the replication. However, the approaches of this kind
are centralised---in another word, even though the agents
in the \MAS may be distributed across the network, the algorithm
itself resides on a single node. Therefore, if the node hosting
the algorithm goes down, then the whole agent system is
compromised. Moreover, most of approaches are static---once
agent replication is done, they forget the survivability issue
afterwards. It is desirable to continuously monitor how well the
MAS is surviving and to make the replication respond to the
changes in the environment. Furthermore, we need efficient
algorithms to measure the performance of the replication so as to
guide the replication method.
We have developed distributed and adaptive architectures and algorithms to ensure the maximal survivability of a MAS. We also proposed various approximations to efficiently compute the survivability.
We implemented and tested different models/algorithms and reported on the advantages and disadvantages of them in different environmental settings.
It is a collaborative project with Prof. Thomas Eiter's group at Vienna. We investigated a major problem that may arise during the implementation of a multiagent system - to verify
that agents are coded correctly and they collaborate well to reach the goal.
It is well known that verification is impossible in general
especially when details of the agents are missing. An alternative
approach is through extensive testing, where large number of
scenarios are created and simulated by the developed multiagent
system in order to ensure the agents are bug-free and with
appropriate behaviours. Clearly, such testing approach is
expensive and inefficient. Debugging tools designed specifically
for multiagent systems are highly desirable.
(click to visit the project's homepage)