Evaluating the Impact of AI-Based Priced Parking with Social Simulation

Abstract

Across the world, increasing numbers of cars in urban centers lead to congestion and adverse effects on public health as well as municipal climate goals. Reflecting cities’ ambitions to mitigate these issues, a growing body of research evaluates the use of innovative pricing strategies for parking, such as Dynamic Pricing (DP), to efficiently manage parking supply and demand. We contribute to this research by exploring the effects of Reinforcement Learning (RL)-based DP on urban parking. In particular, we introduce a theoretical framework for AI-based priced parking under traffic and social constraints. Furthermore, we present a portable and generalizable Agent-Based Model (ABM) for priced parking to evaluate our approach and run extensive experiments comparing the effect of several learners on different urban policy goals. We find that (1) outcomes are highly sensitive to the employed reward function; (2) trade-offs between different goals are challenging to balance; (3) single-reward functions may have unintended consequences on other policy areas (e.g., optimizing occupancy punishes low-income individuals disproportionately). In summary, our observations stress that fair DP schemes need to account for social policy measures beyond traffic parameters such as occupancy or traffic flow.

Appendices

Methodology

Table 1 lists the information contained in $S_t$, and Fig. 4 shows four out of the five deployed reward functions. Figures 5 and 6 visualize the hyperparameter tuning processes that resulted in the utilized model configurations for the ABM.

Table 1. Variables contained in state $S_t$.

Fig. 4.

Reward functions for learning optimal parking prices.

Fig. 5.

Result of Bayesian hyperparameter tuning for 100 episodes for the evaluation model. The target function optimizes for the desired volume of traffic and average share of cruising cars of 30% (consistent with Shoup [11]) (higher is better).

Fig. 6.

Result of random hyperparameter tuning for 100 episodes for the training model. The target function minimizes the euclidean distance to the evaluation configuration in terms of the dimensions described in Sect. 3 (lower is better).

Parameters for Reproduction

Table 2 contains the parameters of our Netlogo ABM, and Table 3 lists the final hyperparameters across all reward functions.¹⁰

Table 2. Model parameters of our NetLogo ABM.

Table 3. Final hyperparameters across all reward functions.

Detailed Results

Figs. 7, 8, 9, 10, 11, 12 and 13 show in-depth results with CPZ fees, occupancy, number of cars, social composition, and car speed per reward function against the time of day.

Fig. 7.

Results of the median static baseline run. (Color figure online)

Fig. 8.

Results of the median dynamic baseline run. (Color figure online)

Fig. 9.

Results of the median $\mathrm {r_{occupancy}}$ run. (Color figure online)

Fig. 10.

Results of the median $\mathrm {r_{cars}}$ run. (Color figure online)

Fig. 11.

Results of the median $\mathrm {r_{social}}$ run. (Color figure online)

Fig. 12.

Results of the median $\mathrm {r_{speed}}$ run. (Color figure online)

Fig. 13.

Results of the median $\mathrm {r_{composite}}$ run. (Color figure online)