Abstract:Policy Mirror Descent (PMD) is a powerful and theoretically sound methodology for sequential decision-making. However, it is not directly applicable to Reinforcement Learning (RL) due to the inaccessibility of explicit action-value functions. We address this challenge by introducing a novel approach based on learning a world model of the environment using conditional mean embeddings. Leveraging tools from operator theory we derive a closed-form expression of the action-value function in terms of the world model via simple matrix operations. Combining these estimators with PMD leads to POWR, a new RL algorithm for which we prove convergence rates to the global optimum. Preliminary experiments in finite and infinite state settings support the effectiveness of our method
What problem does this paper attempt to address?
The main problem that this paper attempts to solve is how to effectively apply the Policy Mirror Descent (PMD) method in Reinforcement Learning (RL). Specifically, the PMD method originally needs to explicitly know the action - value functions of all policies, but this information is usually unavailable in the RL environment. Therefore, the paper proposes a new method of learning the environmental world model based on Conditional Mean Embeddings (CMEs) to estimate the action - value functions, and combines the PMD framework to propose a new RL algorithm - POWR (Policy mirror descent with Operator World - models for Reinforcement learning).
### Main problems and solutions
1. **The applicability of PMD in RL**:
- **Problem**: The PMD method requires explicit knowledge of the action - value functions of all policies. However, in the RL environment, it is impossible to directly obtain this information because the environmental dynamics and reward functions are usually unknown.
- **Solution**: By introducing the world model based on CMEs, the action - value functions can be estimated indirectly. CMEs allow estimating the expected value without a large number of samplings, thus avoiding the sampling error and model error in traditional methods.
2. **The estimation of action - value functions**:
- **Problem**: Traditional methods of estimating action - value functions rely on a large number of samplings, which is unrealistic in practical applications, especially when the state space is large or infinite.
- **Solution**: Using CMEs and operator theory, the paper derives a closed - form expression of the action - value function, which can be efficiently calculated only by matrix operations. This method not only reduces the computational complexity but also improves the accuracy of the estimate.
3. **The extension problem in infinite state space**:
- **Problem**: Traditional methods usually assume that the state space is finite, while in practical applications, the state space may be infinite.
- **Solution**: By extending the PMD framework to infinite state space and combining the learning method of CMEs, the paper proves the convergence of the POWR algorithm in infinite state space and provides a theoretical convergence rate guarantee.
### Main contributions of the paper
1. **Proposes a world model framework based on CMEs**, which can generate estimates of the action - value functions of policies through matrix operations.
2. **Combines the CMEs world model with the mirror descent update step** and proposes an improved PMD algorithm (POWR).
3. **Proves that under an appropriately selected Sobolev space**, the POWR algorithm is well - defined and shows the theoretical result of its convergence to the global optimal solution.
4. **Provides experimental verification** and tests the effectiveness of the POWR algorithm in simple environments with finite and infinite state spaces.
Through these contributions, the paper successfully solves the key problems in the application of the PMD method in RL and provides new directions and tools for future research.