Rethinking Recurrent Neural Networks for Time Series Forecasting: A Reinforced Recurrent Encoder with Prediction-Oriented Proximal Policy Optimization

TL;DR

This paper targets improving encoder-only RNN time series predictors by introducing a Reinforced Recurrent Encoder (RRE) coupled with a Prediction-oriented PPO (PPO4Pred). The approach formulates adaptive input feature selection, dynamic hidden skip connections, and selective output targets as a Markov Decision Process, and optimizes them with a Transformer-based policy and dynamic transition sampling in a co-evolutionary loop. Empirical results across five real-world datasets show substantial gains over static, heuristic, and RL baselines, and even competitive performance with state-of-the-art Transformer models on several horizons. The work demonstrates that carefully learned architectural adaptation within encoder-only RNNs can yield highly accurate and stable multi-step forecasts with practical implications for engineering informatics, albeit at increased training cost that can be mitigated with future efficiency techniques.

Abstract

Time series forecasting plays a crucial role in contemporary engineering information systems for supporting decision-making across various industries, where Recurrent Neural Networks (RNNs) have been widely adopted due to their capability in modeling sequential data. Conventional RNN-based predictors adopt an encoder-only strategy with sliding historical windows as inputs to forecast future values. However, this approach treats all time steps and hidden states equally without considering their distinct contributions to forecasting, leading to suboptimal performance. To address this limitation, we propose a novel Reinforced Recurrent Encoder with Prediction-oriented Proximal Policy Optimization, RRE-PPO4Pred, which significantly improves time series modeling capacity and forecasting accuracy of the RNN models. The core innovations of this method are: (1) A novel Reinforced Recurrent Encoder (RRE) framework that enhances RNNs by formulating their internal adaptation as a Markov Decision Process, creating a unified decision environment capable of learning input feature selection, hidden skip connection, and output target selection; (2) An improved Prediction-oriented Proximal Policy Optimization algorithm, termed PPO4Pred, which is equipped with a Transformer-based agent for temporal reasoning and develops a dynamic transition sampling strategy to enhance sampling efficiency; (3) A co-evolutionary optimization paradigm to facilitate the learning of the RNN predictor and the policy agent, providing adaptive and interactive time series modeling. Comprehensive evaluations on five real-world datasets indicate that our method consistently outperforms existing baselines, and attains accuracy better than state-of-the-art Transformer models, thus providing an advanced time series predictor in engineering informatics.

Figures (18)

• Figure 1: Two RNN architectures for time series forecasting tasks. Solid arrows represent hidden state transitions, and dashed boxes indicate input-output pairs. (a) The Seq2Seq model forms a single training pair (blue dashed box). (b) The Encoder-only model generates multiple training pairs (colored dashed boxes) via an autoregressive structure.
• Figure 2: Illustration of time series forecasting examples in real-world traffic data nie2022time. Historical observations (96 time steps, blue line) are used to generate future forecasts (12-step horizon, orange and green line), where example 2 is obtained by shifting example 1. The dashed boxes in the historical region highlight segments with patterns similar to the future forecasts.
• Figure 3: Architecture of the RRE-PPO4Pred method. Part I: RRE framework with sequential decision-making process; Part II: PPO4Pred algorithm with Transformer-based agent and dynamic transition sampling.
• Figure 4: Illustration of dynamic hidden skip connection with three operational cases. Case 1 (Non-skip, $k_t=0$): Standard RNN computation without skip connections. Case 2 (Skip with sufficient history, $k_t \neq 0$, $t-K>1$): Dynamic skip connection selects from full candidate set $\mathcal{H}_t$ of historical hidden states. Case 3 (Skip with insufficient history, $k_t \neq 0$, $t-K \leq 1$): Early time steps require zero-padding (gray blocks) in candidate set due to insufficient historical states.
• Figure 5: Illustration of asynchronous training and inference.