Predicting the outcome of Automated Negotiations using Machine Learning
Summary
Knowing the outcome of an automated negotiation before it is terminated offers
many advantages. It allows one to change their strategy to obtain a better result, it
gives (multi-issue) negotiators insight into which negotiations are worth the effort
and resources and which are better to terminate, and it can aid in the development of negotiating assistants. Despite these advantages, no research has yet been
done on predicting negotiation outcomes. This thesis tackles the challenge of predicting the outcome of automated negotiations. We limit the scope of these predictions to bilateral alternating-offer negotiations between two time-dependent
agents.
Our method divides the prediction into two parts. First, the bids made by the negotiating agents are converted into utility time series. The future trajectory is then
forecast by six time-series forecasting methods. Second, these forecasts are used
to find a distribution of the most likely intersection points and, with that, a distribution of the most likely negotiation outcomes. In addition, the network gives the
probability of ending the negotiation with an agreement.
In general, neural networks performed best in both time series forecasting and
outcome prediction, achieving an F1-score of 0.876. We found that negotiation
time series are difficult to predict for classic statistical models, as the series have
a low level of predictability. The predictability of time series can be improved by
applying a behavior-based strategy, which we model by introducing a learning
rate.
Our work shows that neural networks are a promising direction for automated
negotiation outcome prediction. We believe these results are only the beginning
of their capabilities and can further be improved by experimenting with more
network architectures and negotiation settings. Therefore, we encourage other
researchers to take the next steps to improve our results.