CHIME: New Method to Improve Long Term Planning Skills of Artificial Intelligence Agents
The complex tasks of artificial intelligence agents are step-by-step solving, making them available in real world applications. However, existing methods require high calculation costs or expensive training data. The researchers developed an innovative system called CHIME (Credit-Aware Hierarchical Memory Evolution) to solve this problem.
The main features of CHIME can be summarized as follows:
- Automatic development during the removal of agent capabilities by storing past experiences in the external memory bank
- Only costs are reduced by learning from the interaction results of the agent without parameter update
- Smartly determined which memories will be held in the forefront thanks to the design of credit-fare
- Haryriarchy structure allows for the separation and management of complex tasks to lower targets
- Traditional test-time offers a much more economical solution compared to search and educational-based methods
This approach allows agents to act faster and effective when they encounter similar new problems by teaching from previous successful task solutions. The memory system not only remembers correct decisions, but also can distinguish which information is really valuable.
Systems such as CHIME in areas such as automation and robotics of long-term tasks can be the key to practicalize artificial intelligence applications. This study published in ArXiv shows that it is possible to keep system costs at minimum while increasing the continuous learning capacity of agents.