Abstract:While fingerprinting localization is favored for its effectiveness, it is hindered by high data acquisition costs and the inaccuracy of static database-based estimates. Addressing these issues, this letter presents an innovative indoor localization method using a data-efficient meta-learning algorithm. This approach, grounded in the ``Learning to Learn'' paradigm of meta-learning, utilizes historical localization tasks to improve adaptability and learning efficiency in dynamic indoor environments. We introduce a task-weighted loss to enhance knowledge transfer within this framework. Our comprehensive experiments confirm the method's robustness and superiority over current benchmarks, achieving a notable 23.13\% average gain in Mean Euclidean Distance, particularly effective in scenarios with limited CSI data.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is that traditional fingerprint positioning methods face the problems of high data acquisition cost and inaccurate estimation based on static databases in dynamic indoor environments. Specifically:
1. **High data acquisition cost**: Traditional fingerprint positioning methods require a large amount of data collection, which is not only time - and resource - consuming, but also difficult to update frequently in practical applications.
2. **Inaccurate estimation based on static databases**: Due to the dynamic changes in the indoor environment (such as the movement of people, the change of furniture layout, etc.), traditional methods based on static databases are difficult to maintain high - precision positioning.
To solve these problems, the paper proposes an innovative indoor positioning method based on few - shot meta - learning. This method uses the knowledge in historical positioning tasks to improve the adaptability and learning efficiency of the model in new environments, thereby achieving more efficient data utilization and lower data acquisition costs.
### Main contributions
1. **Proposed an indoor positioning method based on few - shot meta - learning**: By designing a flexible model, this method can quickly adapt to new environments and use the training knowledge in multiple datasets to enhance data efficiency.
2. **Introduced task - weighted loss**: In order to guide knowledge transfer, especially in a large number of training tasks required by meta - learning algorithms, a task - weighted loss function is proposed.
3. **Experimental verification of the effectiveness and robustness of the method**: Through extensive experiments, the superior performance of this method in different scenarios has been proved, especially in the case of limited CSI data, the average positioning error is significantly reduced, achieving a 23.13% improvement.
### Formula summary
- **Probability representation of task distribution**:
\[
p(T)=\sum_{n_{\text{sub}} = 0}^{N_{\text{sub}}- 1}\sum_{f\in F}p(T|n_{\text{sub}}, f)\cdot p(n_{\text{sub}})\cdot p(f)
\]
where \(p(T|n_{\text{sub}}, f)\) represents the task probability distribution in a specific area and pose, and \(p(n_{\text{sub}})\) and \(p(f)\) represent the prior probabilities of a specific area and pose respectively.
- **Loss function**:
\[
L_T(f_\theta, D)=\frac{1}{KN}\sum_{i = 1}^{KN}(L_i-\hat{L}_i)^2
\]
where \(D\) is the labeled data set of a given task, \(KN\) is the total number of samples in the support set or the query set, and \(L_i\) and \(\hat{L}_i\) are the actual and predicted coordinates respectively.
- **Wasserstein distance**:
\[
W(p, q)=\inf_{\gamma\sim Q(p, q)}\mathbb{E}_{x, y\sim\gamma}[\|x - y\|]
\]
where \(Q(p, q)\) represents the set of all possible joint distributions combining distributions \(p\) and \(q\).
Through these improvements, the method proposed in the paper shows better adaptability and higher positioning accuracy in dynamic indoor environments.