The thesis investigates the use of machine learning as an effective means of supporting autonomous flexibility within complex sensor network management systems. Policy-based management has often been the tool of choice for addressing such requirements, but is often only a partial solution, due to its reliance on end-user capacity for timely and accurate policy creation. A new systems architecture HYBRID, capable of autonomous system flexibility through user-independent adaptation, is therefore proposed. HYBRID combines policy based management with self-learning algorithms to realise a single architecture capable of flexible automation at all levels of a management system. The work described in this thesis demonstrates the limitations of policy-based management, and illustrates how best to mitigate them through the adoption of self-learning techniques. The availability and suitability of today's learning algorithms for facilitating such automation is investigated, and where necessary, algorithmic enhancements to selected techniques are proposed and evaluated to explore relevant complexities. Validity of the architecture is demonstrated through two real -world trials. HYBRID has been applied to address distinct management problems, demonstrating on each occasion, 'how' the proposed architecture supports effective and safe exploitation of machine learning to enable greater behavioural flexibility within complex management systems.