Hierarchical reinforcement learning recommendation system based on dynamic recursion mechanism

Abstract

The invention discloses a hierarchical reinforcement learning recommendation system based on a dynamic recursion mechanism. The system comprises a user portrait corrector which employs a strategy gradient method of the dynamic recursion mechanism and introduces a parameter dynamic sparse weight to delete noise data and modify a user portrait, wherein the strategy gradient method of the dynamic recursion mechanism comprises a dynamic baseline and recursion reinforcement learning based on a time sequence context, the dynamic baseline being used for improvement of a learning strategy for total income by adopting a dynamic sparse weight; an attention mechanism used for automatically adjusting changes of user preferences; and a recommendation model used for recommending the most relevant articles to the user through the attention mechanism. According to the recommendation system, a parameter dynamic sparse weight is introduced into the strategy gradient method, so that an intelligent agentselects an optimal behavior under a global optimal strategy; and secondly, in combination with hierarchical reinforcement learning of the time context, the method can more reliably converge, so that the stability of model prediction is improved.

Description

Technical field[0001]The present invention relates to intelligent recommendation techniques based on hierarchical strengthening, and more particularly to a recommended system based on dynamic recursive mechanism.Background technique[0002]While the hierarchical reinforcement learning, HRLs have made significant progress in the recommendation system, their forecast instability is a key defect, mainly due to the performance of strengthening learning. On the one hand, due to the randomness of the strategy, there are several possible random behavior in each state. For example, the action selected under this policy is random, causing the user portrait that cannot be modified well. On the other hand, the state transition probability is also random (ie, there are multiple possible next states per current state), which makes the expected income from the environment uncertain. Therefore, these two random properties will affect the performance of the recommended system.Inventive content[0003]I...

AI Technical Summary

Problems solved by technology

For example, the actions selected under this policy are random, resulting in not always being able to modify user portraits well

On the other hand, state transition probabilities are also stochastic (i.e. each current state has multiple possible next states), which makes the expected payoff from the environment uncertain

Therefore, both kinds of randomness can affect the performance of recommender systems

Images