Autonomous systems interacting and collaborating need to understand their individual and collective goals as well as their individual capabilities. This keynote outlined challenges and approaches to enable systems to collaborate. Additionally we explore approaches to enable systems to learn and exploit the diverse knowledge of the individual agents in a collaborative manner. The rise of computing systems is unbroken, spreading from our wrist, pockets, to our cars, houses, and even cities [1]. Individual smart devices are being deployed and distributed around our environment with their very individual goals, resources, and capabilities [2, 3]. Systems therefore collaborate only on accident or if they have explicitly been designed to do so [4]. This leads to many systems operating with limited performance as they interfere with each other in the environment. The potential of collaboration with initially unknown systems, that might be able to support each other, is completely untapped. In previous work we relied on common implementations and knowledge when utilising nature-inspired approaches. In contrast, we aim to close the gap and enable systems to acknowledge each other and build upon their individual knowledge and capabilities in more recent and ongoing work.
communication efort required systems to have knowledge about others in their respective
environments in order to target their advertisement [
[
To overcome collaborative challenges, we recently started to explore potentials of collaborative multi-agent learning. Here we explore various strands of research around deep learning and aim to bring those strands together [14]. Specifically, we explore federated learning together with early exits and split computing. Federated learning are techniques where the knowledge of individual learners can be combined in a larger model [15]. By sharing the trained models only, rather than the actual training input, privacy concerns can be mitigated. This is beneficial when the individual learners are unknown or change at runtime. Usually, the tasks for which the networks are trained are known to all agents. However individual agents could utilise available data to estimate the tasks for which a network has been trained by another agent, allowing one agents to draw conclusion about the inference capabilities of another agent. Overall, trustworthiness and correctness can be verified by the individual agents upon receiving the networks by utilising previous training data before merging received networks with local networks.
Early exits is a concept allowing deep neural networks to not execute all neural layers but stop execution at earlier neural layers either because satisfactory results are achieved or if later layers should be executed at another computing location (e.g. from transferring computation from the edge to the cloud) [16]. We specifically work towards approaches where we utilise early exits in combination with federated learning. While general knowledge is shared with other agents, specific knowledge is retained at the individual agent. By sharing the general knowledge, inference can be improved overall. Specialising the individual agents further increases performance under the assumption of observations remaining the same (or similar) for the diferent agents. In future work, we are interested in collaborative agents, able to request support from others in case their own inference result is unsatisfactory. Instead of transmitting all information or even raw input data, only intermediate results from the early exits are being shared among the agents, reducing the amount of information transmitted while preserving important privacy aspects. [11] L. Esterle, J. N. A. Brown, I think therefore you are: Models for interaction in collectives of self-aware cyber-physical systems, ACM Trans. on Cyber-Physical Systems 4 (2020). [12] A. Diaconescu, K. L. Bellman, L. Esterle, H. Giese, S. Götz, P. Lewis, A. Zisman, Architectures for collective self-aware computing systems, in: Self-Aware Computing Systems, 2017, pp. 191–235. [13] L. Esterle, J. N. Brown, The competence awareness window: Knowing what i can and cannot do, in: Int. Conf. on Autonomic Computing and Self-Organizing Systems, 2020, pp. 62–63. [14] L. Esterle, Deep learning in multiagent systems, in: A. Iosifidis, A. Tefas (Eds.), Deep
Learning for Robot Perception and Cognition, 2022, pp. 435–460. [15] J. Konečny`, H. B. McMahan, F. X. Yu, P. Richtárik, A. T. Suresh, D. Bacon, Federated learning:
Strategies for improving communication eficiency, arXiv preprint arXiv:1610.05492 (2016). [16] S. Teerapittayanon, B. McDanel, H.-T. Kung, Branchynet: Fast inference via early exiting from deep neural networks, in: Int. Conf. on Pattern Recognition, 2016, pp. 2464–2469.