g. ). For example, an intelligent agent with the task to autonomously drive a car considering the traffic signs already proves a high degree of agent autonomy. However, if the agent is confronted with a new traffic sign, the learning of this new circumstance might still need human involvement as the agent might not be able to completely learn by itself . Therefore, the necessary involvement of humans, especially in the thinking layer (= executing backend and learning backend), is of major interest when describing AI and the underlying machine learning models. The degree of autonomy for each step of machine learning can be investigated and may help the autonomy of an agent in terms of the related machine learning tasks.

The choice on this feature in general (simple-reflex vs. learning agent) influences the overall design of the agent as well as the contribution of machine learning. The overview of our resulting framework is depicted in figure 2. In conclusion, in the case of a simple-reflex agent, machine learning takes places as a once-trained model in the execution sublayer. In contrast, it plays a role in the learning sublayer of a learning agent to continuously improve the model in the execution sublayer. This improvement is based on knowledge and feedback, which is derived from the environment via the execution layer. to characterize

3.3. Continuum between human involvement and machine involvement 4. Research priorities learning-enabled artificial intelligence for machineWhen it comes to the executing backend and the learning backend, it is not only of importance if and learning models are how underlying machine updatedbut how much automated the necessary processes are. Every machine learning task involves various process steps, including data source selection, data collection, preprocessing, model building, evaluating, deploying, executing and improving (e.g. , , ). While a discussion of the individual steps is beyond the scope of this paper, the autonomy tasks as an and implementation within the agent is of particular interest in each necessary task of the machine learning lifecycle .

The presented framework of machine learning and its role within intelligent agents is still on a conceptual level. However, given the misunderstandings and ambiguity of the two terms , we see potential for further research with the aim both to clarify the terminology and to map uncharted territory for machine-learning enabled artificial intelligence.