A novel adaptive weight selection algorithm for multi-objective multi-agent reinforcement learning

To solve multi-objective problems, multiple reward signals are often scalarized into a single value and further processed using established single-objective problem solving techniques. While the field of multi-objective optimization has made many advances in applying scalarization techniques to obtain good solution trade-offs, the utility of applying these techniques in the multi-objective multi-agent learning domain has not yet been thoroughly investigated. Agents learn the value of their decisions by linearly scalarizing their reward signals at the local level, while acceptable system wide behaviour results. However, the non-linear relationship between weighting parameters of the scalarization function and the learned policy makes the discovery of system wide trade-offs time consuming. Our first contribution is a thorough analysis of well known scalarization schemes within the multi-objective multi-agent reinforcement learning setup. The analysed approaches intelligently explore the weight-space in order to find a wider range of system trade-offs. In our second contribution, we propose a novel adaptive weight algorithm which interacts with the underlying local multi-objective solvers and allows for a better coverage of the Pareto front. Our third contribution is the experimental validation of our approach by learning bi-objective policies in self-organising smart camera networks. We note that our algorithm (i) explores the objective space faster on many problem instances, (ii) obtained solutions that exhibit a larger hypervolume, while (iii) acquiring a greater spread in the objective space.

Learning to be different:heterogeneity and efficiency in distributed smart camera networks

In this paper we study the self-organising behaviour of smart camera networks which use market-based handover of object tracking responsibilities to achieve an efficient allocation of objects to cameras. Specifically, we compare previously known homogeneous configurations, when all cameras use the same marketing strategy, with heterogeneous configurations, when each camera makes use of its own, possibly different marketing strategy. Our first contribution is to establish that such heterogeneity of marketing strategies can lead to system wide outcomes which are Pareto superior when compared to those possible in homogeneous configurations. However, since the particular configuration required to lead to Pareto efficiency in a given scenario will not be known in advance, our second contribution is to show how online learning of marketing strategies at the individual camera level can lead to high performing heterogeneous configurations from the system point of view, extending the Pareto front when compared to the homogeneous case. Our third contribution is to show that in many cases, the dynamic behaviour resulting from online learning leads to global outcomes which extend the Pareto front even when compared to static heterogeneous configurations. Our evaluation considers results obtained from an open source simulation package as well as data from a network of real cameras. © 2013 IEEE.

Temperature management for heterogeneous multi-core FPGAs using adaptive evolutionary multi-objective approaches

Heterogeneous multi-core FPGAs contain different types of cores, which can improve efficiency when used with an effective online task scheduler. However, it is not easy to find the right cores for tasks when there are multiple objectives or dozens of cores. Inappropriate scheduling may cause hot spots which decrease the reliability of the chip. Given that, our research builds a simulating platform to evaluate all kinds of scheduling algorithms on a variety of architectures. On this platform, we provide an online scheduler which uses multi-objective evolutionary algorithm (EA). Comparing the EA and current algorithms such as Predictive Dynamic Thermal Management (PDTM) and Adaptive Temperature Threshold Dynamic Thermal Management (ATDTM), we find some drawbacks in previous work. First, current algorithms are overly dependent on manually set constant parameters. Second, those algorithms neglect optimization for heterogeneous architectures. Third, they use single-objective methods, or use linear weighting method to convert a multi-objective optimization into a single-objective optimization. Unlike other algorithms, the EA is adaptive and does not require resetting parameters when workloads switch from one to another. EAs also improve performance when used on heterogeneous architecture. A efficient Pareto front can be obtained with EAs for the purpose of multiple objectives.