Problems in reinforcement learning training

The effect of matlab reinforcement learning in the training process is better, but the reason for the poor effect after saving the agent is, or how to save the good effect in the training process

[Shantanu Dixit]

Hi Ye, can you elaborate on the issue?

[ye]

That is, in the training process will encounter a good effect of the agent, this time to stop training and save the agent, but with the saved agent to run, the effect and training process is very different

[Shantanu Dixit]

Assuming you're experiencing different training process before and after loading the saved agent, this could be due to following factors:

1. Experience Buffer: By default, the experience buffer isn't saved with some agents like DDPG and DQN. If you plan to continue training the saved agent, consider setting 'SaveExperienceBufferWithAgent' to true to preserve the experience buffer.
2. Non-Determinism and Exploration Strategy: Training may involve stochastic elements, causing the agent to explore different trajectories after being reloaded, which could result in a different training process.

Additionaly you can refer to 'SaveAgentCriteria' and 'SaveAgentValue' to save agents that meet specific performance criteria.

Refer to the below MathWorks documentation for different saving strategies:

Saving Agents: www.mathworks.com/help/reinforcement-learning/ug/train-reinforcement-learning-agents.html