System Architecture
Autonomous Self-determined Plan Generation and Preferences
The participation of an agent in EPOS requires the scheduling of its operations, for instance planning the allocation of some resources, e.g. residential energy consumption. A schedule is referred to as a plan and each agent autonomously and locally generates a finite number of possible plans, meaning that the agent can potentially select and execute any of these. The possible plans represent the agents’ flexibility and they may be equivalent for agents or agents may have preferences towards them.
The possible plans represent the agents’ flexibility and they may be equivalent for agents or agents may have preferences towards them.
These preferences can be measured by a local cost function. Given the local cost of the selected plan in each agent, the system has an average local cost indicating whether agents choose preferred plans. Given the local costs, the system fairness can be measured via a dispersion metric, e.g. standard deviation, of the average local costs among the agents.
Collective Plan Selection for Sustainability
(ii) Agents can express their individual preferences for a possible plan via a weighting scheme over the evaluated possible plans. (iii) The dispersion of the cost in the selected plans among the agents is an indicator of fairness. Measurable trade-offs between performance vs. fairness can be made.
Measurable trade-offs between performance vs. fairness can be made.
(iv) Selections are collective by remotely exchanging between agents local and aggregated information about the agent plans. All these aspects are encoded in a fitness function that each agent uses to make its plan selection.
Selections are collective by remotely exchanging between agents local and aggregated information about the agent plans.
Hierarchically Structured Interactions over Tree Topologies
Agents in EPOS are self-organized in a tree topology over which they perform collective decision-making. A tree is a connected acyclic graph over which a duplicate-free aggregation can be performed as required for the execution of the agents operations. By using a tree, an EPOS execution efficiently completes via a bottom-up and top-down exchange of messages.
Agents in EPOS are self-organized in a tree topology over which they perform collective decision-making.
Agents can be sorted in the tree to improve the performance of the global responses. Criteria for sorting agents may include measurable features of the plans, for instance, the variance, or the network availability so that agent disconnections from the tree have a minimal impact on the overall topology. Building and maintenance of tree topologies can be achieved via self-organization using several fully decentralized services, such as AETOS, the Adaptive Epidemic Tree Overlay Service.
