Optical Burst Switching (OBS) is widely believed to be the technology for the future core network in the Internet. Burst assembly time at the ingress node is known to affect the traffic characteristics and loss distribution in the core network. We propose an algorithm for adapting the burst assembly time based on the observed loss pattern in the network. The proposed Learning-based Burst Assembly (LBA) algorithm uses learning automata which probe the loss in the network periodically and change the assembly time at the ingress node to a favorable one. We use a discrete set of values for the burst assembly time that can be selected and assign a probability to each of them. The probability of selecting an assembly time is updated depending on the loss measured over the path using a Linear Reward-Penalty (LR-P) scheme. The convergence of these probabilities eventually leads to the selection of an optimal burst assembly time that minimizes the burst loss probability (BLP) for any given traffic pattern. We present simulation results for different types of traffic and two network topologies to demonstrate that LBA achieves lower BLP compared to the fixed and adaptive burst assembly mechanisms existing in the literature. © IFIP International Federation for Information Processing 2007.