Learning Agents

Solving optimization problems for scheduling tasks can be very complex. Bigger problems can easily need exponential calculation time relative to the size of the problem. Still, some of these problems, like the traveling salesman problem, are relatively easy to grasp for people. This type of problem can provide an easy introduction for researchers or students interested in the subject of AI, particularly multi-agent systems. The AbstractSwarm Multi-Agent Logistics Competition supplies this opportunity. Participants can create and test their own agents on a priori unknown logistics problems. I used it as an entrance to this field in my bachelor’s thesis. Three different multi-agents were implemented: AgentPlace, AgentLearning, and AgentProbabilities. All of these agents share a foundational logic, first identifying a problem as either directed or undirected. The second agent builds upon this by choosing from a set of different heuristics to weight the problem attributes in undirected scenarios. AgentProbabilities introduces a meta-learning layer, dynamically adjusting the selection probability for each heuristic based on its observed success. The resulting implementations have won the competition in 2022 and unofficially in 2023.