TY  - JOUR
T1  - Curiosity-Driven TD-MPC with Model-Free Reinfocement Learning Policy
AU  - Ji, Chang-Hun 
AU  - Kim, Ju-Bong 
AU  - Han, Youn-Hee 
JO  - The Journal of Korean Institute of Communications and Information Sciences
PY  - 2023
DA  - 2023/1/14
DO  - 10.7840/kics.2023.48.6.712
KW  - Deep Reinforcement Learning
KW  - Model-based Reinforcement Learning
KW  - Model Predictive Control
KW  - Precise Control
AB  - TD-MPC, which has the highest performance among recent model-based reinforcement learning algorithms, extracts behaviors from model predictive control and DDPG agents in the learning process. However, the DDPG agent does not apply the extracted behavior to the environment, but only applies the behavior extracted from model predicted control to the environment. In this paper, we propose an enhanced TD-MPC that utilizes a dual policy that applies to the environment by considering both the DDPG agent and model predictive control of TD-MPC. In addition, by encouraging exploration based on curiosity, bias in utilization that can occur when choosing an action between dual policies is addressed. It is confirmed that the algorithm proposed in various environments of the DeepMind Control Suite has higher sample efficiency and higher performance than the existing TD-MPC.