Neuro-symbolic model for real-time forecasting problems

The results of an investigation are presented into the development of a strategy for Cooperative problem solving.  The work focuses on the possible combinations of case-based reasoning and artificial neural networks and their integration. The results obtained demonstrate that for a particular type of problem, a case based reasoning system assisted by an artificial neural network, forming a hybrid artificial intelligence system, can provide a more effective problem solving capability than either of the methods by themselves.

In investigating its effectiveness in a practical real world situation, the approach is applied to a particular problem in the field of oceanography: the prediction of the structure of the temperature of the water ahead of an ongoing vessel in real time. The complexity of the application domain is due to the high activity of the oceanic water masses and to the diversity of their characteristics. A number of algorithms have been investigated in recent years, with the aim of providing a model capable of predicting the value of the ocean temperature several steps ahead of a thermal time series, but without producing sufficiently accurate results.

The hybrid model that has been developed integrates, in a single problem solving mechanism, cycles of an artificial neural network and case based reasoning processes.  The hybrid forecasting model has been successfully tested on a cruise in real time, and the results obtained show that this model provided more accurate results than either the case-based reasoning model or the artificial neural networks by themselves. The hybrid model has demonstrated the potential for solving problems, in real time, in situations where it is difficult to define the rules that determine the behaviour of the parameters that may be forecasted.

The success of the system in generating effective forecasts may be attributed to the combination of an extensive database of past cases, supported by the neural adaptation mechanism which each time around the forecasting cycle, enables the forecasting process to learn from all the selected closely matching cases.

The experimental results obtained are encouraging and indicate that the neural network supported approach is effective in the task of predicting future oceanographic parameter values. Extrapolating beyond these results it is believed that the approach may be applicable to the problem of parametric forecasting in other complex domains using sampled time series data.