Title:
An Architecture for Purely Probabilistic Negotiating Agents:  
Pessimism and Punishment, Laissez-Faire Paths, and One-Sided Rationality

Ronald P. Loui
Department of Computer Science and Engineering
Washington University
St. Louis
USA

Abstract

This talk describes a major new model of negotiation and a major new
architecture for automated negotiating agents.  The agents are driven
purely by the expected utility of continuing negotiation based on
objective probability of breakdown.  This internal force is identified as
pessimism.  The model requires a compensatory force, a kind of resentment,
which increases the agent's willingness to punish negotiating partners for
procedural offenses.  Pessimism requires a theory of induction.
Punishment requires nonstandard process utility attached to the act of
unilateral breakdown.

The resulting picture of negotiation is intended to displace the solution
concept in game-theoretic models of negotiation.  The new concepts are
control-theoretic:  path, state, process, progress, admissibility,
controllability, and estimation.  We aim to avoid several problems with
game-theoretic approaches to negotiation; specifically, we exclude Nash
equilibrium, recursive modeling of beliefs, and path-independence.  We
focus on procedural fairness rather than the recently celebrated
substantive distributive fairness (because the latter is modeled by a
trivial transformation on the payoff matrix).

In the P&P model, agents may produce proposals ahead of schedule,
especially because the laissez-faire path leads to breakdown.  Agents may
also have memory of resentment which is useful in preventing manipulation
through accurate estimation of parameters.

In honor of Henry Kyburg, we note the relationships between interval
probability and one-sided rational constraint.  We note that the model
would not be possible without a foundation of empirical, epistemological,
objective probability (Kyburg 61, 74).  We note how a non-Bayesian,
scientific-theory-formation approach to utility theory is a prerequisite
to nonstandard process utilities (Kyburg 84, 90).  We note that the model
arises simply by asking what would have been the most straightforward,
most logical approach to modeling negotiation, in 1950, if probabilists
had had the greater subtlety and sophistication made possible by a
generation of great inductivists and decision theorists.

Simulation results will be reported, and interactive simulators will be
available for demonstration.  This author also suggests broader ways that
AI is producing lasting new models of negotiation, including the modeling
of cheap talk, heuristic valuation of payoffs, and explicit joint
problem-solving.