Preliminary experiments with learning agents in an interactive multi-agent systems architecture tradespace exploration tool

One of the main problems of current system design and architecture tools is the communication between the tool and the user and vice-versa. The tools are too opaque, and the users often don´t understand their outputs, which results in loss of confidence in the tool. The users are also frustrated because they don´t have an effective way of providing their expert knowledge (acquired either in real time or a priori) to the tools, which results in poor performance. This work describes an on-going effort to improve that communication by incorporating an interactive and “transparent” learning agent into a multi-agent tradespace exploration tool. This learning agent mines the current population of architectures for driving features (combination of architectural variables) that appear to drive architectures towards a “good region” or a “bad region” of the tradespace, and shares that information with the user. This information is then used to produce a surrogate classification model based on a decision tree that predicts whether or not an architecture is likely to be in the “good region”. The decision tree model was chosen based on evidence that it is among the most human understandable classification models. The information about driving features is also fed to an adaptive heuristic optimization agent that uses it to intelligently apply “good features” to or remove “bad features” from existing architectures, with the hope of improving the efficiency of the search process. Perhaps most importantly, information about driving features, surrogate models, and heuristics can help the user understand the results of the tool better and gain useful architectural insight, such as dominant features and trade-offs in the architecture space.