UCT (Upper Confidence Bound for Trees) is a Monte Carlo tree search method that decides at each step which branch of a search tree the search will follow next. UCT applies the „Upper Confidence Bound“ (UCB) principle to search trees and balances two goals: exploration (exploration) and exploitation (exploitation).
Classification: Monte Carlo tree search
The Monte Carlo Baum Search (MCTS) is a search algorithm. It approaches a solution through many random branches. It is used when precise calculation is impossible or very time-consuming, for example in reinforcement learning.
This is how UCT selects a branch
At each node, UCT first selects actions that have never been selected there before. Under these circumstances, chance plays a role. Once all actions have been tried, UCT calculates a value for each action from two parts and selects the action with the highest value.
The first part is the average yield measured so far for the action. It favors branches that have proven themselves. The second part is an exploration bonus. This bonus increases the more often the parent node is visited, without the branch in question being selected. This gives a rarely visited branch a chance once in a while. An exploration parameter weights the exploration bonus and thus controls how often the search tries new branches.
Why compensation matters
Without this compensation, two errors threaten. A method that makes heavy use of heuristics often selects the same action. It then settles too early on a seemingly good, but actually weaker, branch and remains stuck in a local optimum. A method without binding to proven branches, on the other hand, tries endlessly to generate new branches without ever going deep enough.
UCT in AI agents
Language Agent Tree Search (LATS) transfers the tree search to language model agents. In the selection phase, LATS uses the UCT formula. It weighs the previously measured value of a node against the ratio of previous visits, thus determining the node to be searched for next.
Reference to SAP Business One
the Availability check in SAP Business One, several solution options are displayed simultaneously when the Order quantity The available stock exceeds the available stock. However, the system does not evaluate these options and does not automatically select any of them. There are no measured values for each option that a selection rule such as UCT could access. The choice remains entirely with the user.
Demarcation
UCT is not the tree search itself, but a selection rule within the tree search. The Monte Carlo tree search includes further steps, such as attaching new nodes and backpropagation, in which the result of a traversal in the tree is returned upward. UCT only decides which branch is next in line.
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
