Strategyproof Deep Reinforcement Learning for Online Resource Allocation
The neural allocation rule reached 97.8% of an offline welfare upper bound in a four-resource test while guaranteeing truthful reporting.
The mechanism pairs a monotonic neural network with a payment set at the lowest value that would still secure acceptance, holding other inputs fixed. The authors state that this combination ensures strategyproofness and individual rationality. On the tested resource stream, a linear learner reached 93.8% of the welfare bound and first-come, first-served reached 89.0%. Training by imitation reached 97.5% in 261 seconds, about 89 times faster than reinforcement learning.
ArXiv · AI/CL/LG's note
The mechanism pairs a monotonic neural network with a payment set at the lowest value that would still secure acceptance, holding other inputs fixed. The authors state that this combination ensures strategyproofness and individual rationality. On the tested resource stream, a linear learner reached 93.8% of the welfare bound and first-come, first-served reached 89.0%. Training by imitation reached 97.5% in 261 seconds, about 89 times faster than reinforcement learning.
ArXiv · AI/CL/LG's note
score 4