首页>
外国专利>
NEURAL NETWORK MODEL FOR REACHING A GOAL STATE
NEURAL NETWORK MODEL FOR REACHING A GOAL STATE
展开▼
机译:达到目标状态的神经网络模型
展开▼
页面导航
摘要
著录项
相似文献
摘要
An object, such as a robot, is located at an initialstate in a finite state space area and moves under thecontrol of the unsupervised neural network model of theinvention. The network instructs the object to move inone of several directions from the initial state. Uponreaching another state, the model again instructs theobject to move in one of several directions. Theseinstructions continue until either: a) the object hascompleted a cycle by ending up back at a state it hasbeen to previously during this cycle, or b) the objecthas completed a cycle by reaching the goal state. If theobject ends up back at a state it has been to previouslyduring this cycle, the neural network model ends thecycle and immediately begins a new cycle from the presentlocation. When the object reaches the goal state, theneural network model learns that this path is productivetowards reaching the goal state, and is given delayedreinforcement in the form of a "reward". Upon reaching astate, the neural network model calculates a level ofsatisfaction with its progress towards reaching the goalstate. If the level of satisfaction is low, the neuralnetwork model is more likely to override what has beenlearned thus far and deviate from a path known to lead tothe goal state to experiment with new and possibly betterpaths. If the level of satisfaction is high, the neuralnetwork model is much less likely to experiment with newpaths. The object is guaranteed to eventually find thebest path to the goal state from any starting location,assuming that the level of satisfaction does not exceed athreshold point where learning ceases.
展开▼