Load-bearing engineering structures typically have a static shape fixed during design based on expected usage and associated load cases. But neither can all possible loading situations be foreseen, nor could this large set of conditions be reflected in a practical design methodology - and even if either or even both was possible, the result could only be the best compromise and thus deviate significantly from the optimum solution for any specific load case.
A structure that could change its local properties in service based on the identified loading situation could  potentially raise additional weight saving potentials and thus support lightweight design, and in consequence, sustainability.
Property change of materials has been discussed in the past, e.g, in terms of stiffness changes etc. The present paper provides and overview of such approaches. Materials of this kind would exhibit a cellular architecture consisting of a large number of active cells with sensing and actuation capabilities as basis for local change. As additional part of these active cells, suitable control mechanisms both in terms of algorithms and hardware units are necessary. One major issue existing with a fine-grained active smart cellular structure is the correlated control of each actuator and the informational organization. Among the requirements for the control system there are real-time capabilities and high levels of robustness.
As a control mechanism behind property adaptation, a two-stage approach combining mobile & reactive Multi-agent Systems (MAS) and Machine Learning is foreseen. MAS are used to analyze the loading situation based on sensor information - preferably, highly localized strain data - and negotiate a matching redistribution of, e.g., local stiffness values according to some higher-level aim like minimizing total elastic strain energy or maximum stress levels in the system. The machine learning approach in contrast would recognise loading situations that have already been encountered in the past and on this basis avoid the MAS approach by directly proposing the solution found in the preceding case.  
To achieve these aims, the system should feature self-organization and self-adaptivity in terms of computational units. Planning of the agents (i.e., planning of control actions) should base on Distributed agent-based Machine Learning (DML). In this work, a hybrid learning approach is used with prediction of already known load situations (i.e., supervised trained learning) and reinforcement learning to improve the material adaptation by minimizing selected target parameters. This is performed by the agents by adapting their action planning based on the learning results. One major feature of the DML is the deployment of a collection of spatially distributed learner agents, each learning a local model, which are finally fusioned to a global learner model via negotiation, following a fine-grained Divide-and-Conquer approach.
The present study proposes a virtual evaluation system to analyze potential benefits and develop associated algorithms. This proof of concept should be performed by combining FEM and MAS simulation. The FEM simulation is also used for off-line training of the MAS prior to deployment in a real structure. The classification models learned this way are a starting point and can be updated at run-time by using incremental learning techniques.
Besides providing outlines of the system evaluation, the study will discuss further possible benefits of a system of this kind, including e.g. the possibility of isolating internal damage and compensating its effect on structural performance.