Optimization method and device for game agent training, terminal device and storage medium

Abstract

The invention discloses an optimization method and device for game agent training, a terminal device and a storage medium. The method comprises the steps of acquiring a first observation sequence according to a game script, and setting the sequence as a current observation sequence; based on a pre-trained self-coding network, according to the current observation sequence, obtaining a next observation sequence and a current environmental reward; based on the pre-trained self-coding network and a pre-trained prediction network, acquiring a current self-driven reward according to the current observation sequence and the next observation sequence; adding the current environmental reward and the current self-driven reward to a cumulative reward, and determining whether the cumulative reward isless than a target value; if the cumulative reward is less than the target value, keeping training, taking the next observation sequence as the current observation sequence until the cumulative rewardreaches the target value, and stopping training. The method enables a game agent to obtain the additional self-driven reward for training, thereby improving the training efficiency and saving operating resources.

Description

technical field [0001] The present invention relates to the field of artificial intelligence, in particular to an optimization method, device, terminal equipment and storage medium for game agent training. Background technique [0002] Game agents (Game Agents) are non-player individuals that can achieve human-like intelligent behavior through algorithms in the game environment. Training agents in a game environment has become a common method for artificial intelligence research and development. Reinforcement learning, as a commonly used method for training game agents, can enable the trained game agents to complete many complex tasks in the interaction of the game environment; the game agents can learn the gameplay similar to human intelligence by interacting with the game environment. Its goal is to learn the strategy, that is, how to execute each game step to achieve the ideal state. Through reinforcement learning, the game agent uses the set program to conduct self-expl...

AI Technical Summary

Problems solved by technology

[0003] However, in many game environments, the environmental rewards may be very sparse, that is, the game agent needs to take a long series of correct actions to achieve the target reward, and when the environmental rewards are very sparse, it is easy to cause the self-exploration of the game agent to fall into a trap. A large number of repeated invalid attempts lead to low training efficiency and waste of game agent operating resources

Images