Page 272 - Socially Intelligent Agents Creating Relationships with Computers and Robots

Challenges in ABSS of Negotiation 255

though it is unlikely that a wholesale change in rankings will follow from every
failure to achieve some agreement.
In order to capture this learning process about endorsements and their relative
values, agents’ learning is represented by the Chialvo-Bak [1] algorithm.
This algorithm uses a sort of neural network approach but without positive
reinforcement of synapse weights. In the present case, the input neurons are
attached to endorsement tokens and the output neurons are ranking values to
be attached to the endorsements. There were five intermediate layers, each
containing 40 neurons. Starting with the input neurons, each neuron as seven
synapses out to the next layer until the output neuron layer is reached. The paths
followed from input to output neurons is determined by the synapse with the
highest weight emanating from each neuron. When agreement is not reached,
the value of each synapse on the dominant path is reduced by a small amount
(usually by one per cent) and the sum of the reductions is distributed equally
amongthe rest of the (2000+) synapses. Consequently, changes inthe behaviour
of anagent takeplacerelatively infrequentlybut will, from timetotime, befairly
extensive.
There are two advantages to be gained from implementing this learning
process. One is that the simulations determine the most important criteria to
be used in choosing negotiating partners. The other is the flexibility of the
ordering of criteria since it is possible that the importance of different criteria
will change over the course of any negotiation process. It is possible, for
example, that reliability is most important at early stages so that there is some
meaningful communication but that trustworthiness is most important in the
final stages.

2.3 Negotiation strategy

It is a commonplace in the negotiation literature that the least important
issues should be addressed first. Once negotiating styles have accommodated
one another and a recognition of reliability and trustworthiness established,
there is a basis for considering more important substantive issues. The most
difficult issues are left to the last.
Every agent in the model reported here adopts this sort of strategy. Each
agent offers to its preferred negotiating partner a list of positions for the issues
the agent found least important among all of the issues that had not yet been
resolved. Denote the first agent as A and A’s preferred negotiating partner as P.
If P made some offer of negotiating positions then, if that offer contained values
for positions that A found least important, and also some values that A found
to be more important, then A would accept P’s offer on the least important
issues in exchange for P’s acceptance of the same number of A’s positions.
In general terms, some agreement could always be reached provided the two