Temporal Transfer Learning for Traffic Optimization with Coarse-grained Advisory Autonomy

TL;DR

This work tackles traffic optimization with coarse-grained advisory autonomy by using Temporal Transfer Learning (TTL) to transfer knowledge across hold durations $\delta$. It introduces Greedy Temporal Transfer Learning (GTTL) and Coarse-to-fine Temporal Transfer Learning (CTTL) to select source tasks that maximize the aggregate performance $A$ over $[\delta_{min},\delta_{max}]$ under a linear generalization gap with slope $\theta$. Theoretical results show that $A_K^{CTTL}=(1-\frac{1}{4K})\theta(\delta_{max}-\delta_{min})^2$ and provide suboptimality bounds for GTTL relative to CTTL, with empirical evaluation across three mixed-traffic scenarios demonstrating TTL methods outperform baselines and approach Oracle Transfer with a modest number of source tasks. The findings indicate TTL enables data-efficient, human-in-the-loop traffic optimization, offering practical pathways to implement near-term system-level improvements in mixed autonomy settings and suggesting broader applicability to temporally structured RL problems.

Abstract

The recent development of connected and automated vehicle (CAV) technologies has spurred investigations to optimize dense urban traffic to maximize vehicle speed and throughput. This paper explores advisory autonomy, in which real-time driving advisories are issued to the human drivers, thus achieving near-term performance of automated vehicles. Due to the complexity of traffic systems, recent studies of coordinating CAVs have resorted to leveraging deep reinforcement learning (RL). Coarse-grained advisory is formalized as zero-order holds, and we consider a range of hold duration from 0.1 to 40 seconds. However, despite the similarity of the higher frequency tasks on CAVs, a direct application of deep RL fails to be generalized to advisory autonomy tasks. To overcome this, we utilize zero-shot transfer, training policies on a set of source tasks--specific traffic scenarios with designated hold durations--and then evaluating the efficacy of these policies on different target tasks. We introduce Temporal Transfer Learning (TTL) algorithms to select source tasks for zero-shot transfer, systematically leveraging the temporal structure to solve the full range of tasks. TTL selects the most suitable source tasks to maximize the performance of the range of tasks. We validate our algorithms on diverse mixed-traffic scenarios, demonstrating that TTL more reliably solves the tasks than baselines. This paper underscores the potential of coarse-grained advisory autonomy with TTL in traffic flow optimization.

Figures (12)

• Figure 1: Illustrative figure of Temporal Transfer Learning (TTL) for the coarse-grained advisory system. In a coarse-grained advisory system, vehicles receive persistent guidance for a specified hold duration rather than instantaneous controls. The system performance of this system shows the non-robustness to the hold duration of deep reinforcement learning when trained exhaustively. In that, we propose Temporal Transfer Learning (TTL) methods designed to select the source training tasks based on the temporal features. In comparison to the exhaustive and multi-task training methods, TTL gives an intermediate number of policies to train to solve a full set of tasks.
• Figure 2: Two types of advisory system to the human drivers: acceleration guidance (\ref{['fig:guide-accel']}), speed guidance (\ref{['fig:guide-speed']}).
• Figure 3: Visualization of sequential source task selection and corresponding performance evaluations within the guidance hold duration space. The shaded region represents the aggregate performance $A_1$ after selecting $\delta^1$ in the first step. The generalization gap $\Delta J(\delta^1,\delta)$ quantifies the performance drop when applying the policy trained at $\delta^1$ to a target task with $\delta$. At the second step, the selection of $\delta^2$ updates the estimated performance of task with duration of $\delta$ from $J_1(\delta)$ to $J_2(\delta)$.
• Figure 4: An exemplified representation of the Temporal Transfer Learning (TTL) process for source task selection. The graphic showcases the stepwise procedure for two iterations ($k=2$), resulting in two segments demarcated by inflection points at $\delta^1$ and $\delta^2$. The upper-bound performance $J^*$ is indicated by the blue dotted line, as posited in \ref{['assume:constant-upperbound-j']}, while the piecewise linear segments and their slopes, as governed by \ref{['assume:linear-transfer']} and \ref{['assume:same-transfer-slope']}, guide the selection of the next hold duration $\delta^k$ that will maximize the aggregate performance $A_k$. Each segment is assessed for its potential marginal contribution to $A_k$, with decisions influenced by the shape of the performance function $J(\delta)$, here visualized as transitions from the orange to the green area, signifying the shift in guidance hold duration from $\delta^1 = 20$ to $\delta^2 = 33.33$.
• Figure 5: Illustrative figure of Temporal Transfer Learning (TTL) algorithms: Selecting the training task based on the TTL algorithm, evaluating each task based on the trained policies, and taking the best-performing policy for each task.

Theorems & Definitions (13)

• Definition 1: Generalization gap $\Delta J(\delta_S, \delta_T)$
• Definition 2: Sequential Source Tasks Selection Problem
• Theorem 1: Optimal source task selection for greedy transfer
• Definition 3: Cumulative area under the estimated performance function at each iteration
• Lemma 2: Lower bound of Greedy Temporal Transfer Learning
• proof
• Theorem 3: The number of source tasks required to cover the area
• Lemma 4: Optimality of Coarse-to-fine Temporal Transfer Learning
• Theorem 5: Suboptimality of Greedy Temporal Transfer Learning
• proof