Computer Network Defence is concerned with the active protection of information technology infrastructure against malicious and accidental incidents. Given the growing complexity of IT systems and the speed at which automated attacks can be launched, implementing timely and efficient network incident mitigating actions, whether proactive or reactive, is a great challenge. We refer to the automation of action selection and implementation in this domain as Autonomic Computer Network Defence. In this work, we suggest that Autonomic Computer Network Defence can be achieved using Reinforcement Learning and dynamic risk assessment to learn the optimal action sequence, or policy, to recover from given computer network risk situations. Such a policy could then be used by commercial network management and security products to implement selected mitigating actions automatically, as risk states are sensed.