Hybrid Symbolic-Neural Approaches to Artificial Intelligence for Interstellar Missions
Interstellar flight represents an unforgiving environment for autonomous operations with many unknowns imposing the necessity for advanced artificial intelligence (AI). Evidently, interstellar distances preclude any human intervention from Earth. We take the view that the interstellar transit phase of an interstellar mission may be accommodated with a mixture of pre-programmed reactive symbol-manipulation-based intelligence implementing traditional Kalman filter-like algorithms. However, for in-situ exploration at the destination extrasolar system, near-human level capabilities will be required for full autonomy. The implementation of landers and rovers for conducting scientific investigations on planetary surfaces reinforce this requirement for adaptive behaviour. If the interstellar probe is a self-replicating probe, scientific survey is merely the first measurement phase of an in-situ resource utilisation task involving highly complex interactive transformation of planetary environments.
We review and assess state-of-the-art AI to enable robotic machines to perform complex interactive tasks on exoplanetary environments at near human-level competence.
Machine learning methods such as deep learning neural networks are ascendent in AI but they require extensive sets of training data which will be scant. Training systems on solar system missions and deploying them on extrasolar systems transfer learning will not be effective. We have encountered issues in transfer learning due to input-output data pair training inadequacies. One means to circumvent data scarcity is to employ a symbolic knowledge manipulation approach such as expert systems based on Bayesian nets despite their brittleness. Hybridisation represents a suite of approaches that fuses the two paradigms – symbolic knowledge bases with neural networks – to exploit the strengths of both while compensating for their respective weaknesses. There are many different approaches to such hybrid symbolic-neural artificial intelligence that we shall review including Bayesian networks, fuzzy nets, long short term memory, etc. Many involve symbolic methods pre-shaping the neural network into starting configurations to constrain subsequent learning. There are however severe challenges in consistent representational formats.
Our investigations reveal several shortcomings in AI methods but hybridisation presents one approach that may be promising.
Unless these issues are addressed, it is unlikely that we can implement a successful interstellar mission beyond a rapid flyby.
