Backpropagation In tree search, it refers to the step in which the result of a search traversal is moved up in the search tree to update the values of all the nodes on the traversed path. This step is also called „backup“.
Classification in the Monte Carlo tree search
In Monte Carlo Tree Search (MCTS), each search run consists of two phases. First, the search selects nodes from the root based on a search strategy until it reaches a state outside the tree. There it attaches a new node and estimates its value for the first time, often by simulation up to the time horizon. Then follows backpropagation: the output of the run moves up the tree.
What exactly is being updated
In the UCT selection algorithm, the backpropagation process is as follows: For each node on the path, from bottom to root, the visit counter increases. At the same time, the algorithm calculates the average value of the node with the new output as a running average. Other variants of the tree search update other value variables instead of averages. UCT uses the updated counters and averages in the next iteration to select the next branch.
Backpropagation in AI agents
Language Agent Tree Search (LATS) transfers this step to language model agents. LATS first proceeds with the best-rated new branch until success, failure, or the end of the budget is reached. The result of this simulation then updates the ratings of all nodes that have been traversed along the way retroactively. LATS does not adjust model weights in this process.
Reference to SAP Business One
The availability check in SAP Business One shows several solution options simultaneously when the stock is too low. As soon as the user selects one, the system forgets about the case again. There is no backward update that learns from the outcome of a case and influences future displays.
As a thought experiment (no existing function and no announcement), an extension could be considered: the system records how a case turned out and highlights proven options in similar cases. Such a feedback would need a measure of „well-resolved“ cases. This measure would need to be translated into a number so that a system could process it.
Demarcation
Backpropagation in the tree search updates node values and visit counters in the search tree; no model weights are used. The term “backpropagation” in the training of artificial neural networks means something different and is distinct from this step in the tree search.
Sprint and Swirl: How an AI agent divides work and learns from trials
LATS: Considering several paths simultaneously with AI — and the availability check in SAP B1
ReAct and RLEF: How Artificial Intelligence Learns from Real Mistakes — and What Automated Bank Reconciliation in SAP Business One Has to Do with It
SAP B1 10.0 FP2608: Service Layer AI as a transactional layer
Process Reward Models: Why a correct result does not yet prove a correct method
