Deep RL for Combinatorial Optimisation

Session date: 30 September 2024

Session host: Wendelin Böhmer

Summary:

I introduce deep RL and show in which optimization tasks it excels. I will show some of my past (unsuccessful) attempts to learn RL agents that solve combinatorial optimization problems (Knapsack and traveling salesman), and introduce another work that succeeded where I failed. By contrasting these two approaches I hypothesize what the necessary properties (of the problem, algorithm and architecture) are to make deep RL work in this field. I hope to incite a lively discussion which of the audiences problems might have these properties, and/or how we could devise neural architectures that would exhibit them.