Center for Multidisciplinary Research in Traffic
Note: This article is not yet complete đ
This is work that I have done with Dr. Jia Li at Texas Tech University from January 2019 to May 2020. Dr. Li had to leave TTU and the US temporarily during the pandemic and we lost touch so we never got the chance to publish on this work. Nevertheless, find the article below as written during May of 2020.
We investigate spontaneous platoon formation in heterogeneous traffic flow consisting of human-driven (HVs) and autonomous vehicles (AVs) in different behavior scenarios through simulation experiments. Our study reveals that platoons may form spontaneously, and platooning properties are associated with the behaviors of AVs. We conduct the simulation experiments through a parsimonious Cellular Automata model (we develop a software for this purpose), which captures the different characters of AVs and HVs as well as their interactions. AVs are endowed with neighbor awareness and opportunistic behaviors. We observe that, intriguingly, even with this relatively simple model, AVs form into platoons without centralized control. Such phenomena may relate to the intrinsic incentives that AVs perceive and their ability to tell neighbor vehicle types. Our findings indicate the potential of regulating future heterogeneous traffic flow through decentralized agent behavior design.
Driver-less cars (i.e., autonomous vehicles, henceforth called AVs) have already appeared on the road and shared right of way with human-driven vehicles (henceforth called HVs). An appealing feature of AVs is their capability to form platoons, which is anticipated to boost efficiency of traffic flow substantially. A number of recent research have delved into modeling and operations of the mixed flow of AVs and HVs, most of them attempted to capture or optimize the platooning feature of AVs.
Much attention in literature (see e.g.
The key challenge to model mixed flow dynamics of AVs and HVs lies in capturing agentsâ longitudinal and lateral behaviors simultaneously, which should account for not only agentsâ dynamical characters (e.g. their longitudinal speed and acceleration as functions of headway), but also agentsâ decision-making, as well as how the decisions of individual agents aggregate. Such interactions in mixed flow of AVs and HVs is conceivably more complicated than mixed human-driving traffic due to the rich possibilities of AV behaviors. For instance, with the advanced sensing and communication capabilities, AVs may behave more opportunistically than HVs in seeking peers in its neighborhood, because this potentially allow them to form platoons with peers and benefit from more smooth driving and better traveling speed.
In this paper, we are interested in understanding the collective dynamics of AV-HV flow, as a result of the interactions between heterogeneous agents. In particular, we seek to answer the following research questions: when AVs are self-interested and individually controlled by themselves (as opposed to coordinated by system operator), will they be able to self-organize into platoons? And if so, what are the conditions for the self-organization to initiate and sustain? This research is partly inspired from the self-organized phenomena widely existing in biological systems, such as army ants
To answer the above questions, we propose a Cellular Automaton (CA) model to encapsulate the essential behavioral factors of AVs and HVs, and conduct simulation experiments to characterize the impact of these behavior factors on the formation of clusters in mixed traffic flow. The simulation experiments were conducted on a circular road with three lanes. We consider two alternatives of AV behaviors. In the first scenario, AVs behave like HVs, except the difference in free flow speed and braking probability. In the second scenario, AVs are assumed to be fully aware of vehicle types in neighborhood and opportunistic in gap seeking.
The major finding of this research is the self-organized formation of AV clusters and lanes without any centralized control. This finding may suggest clustering as an intrinsic property of mixed flow, when AVs and HVs interact, and AVs are endowed with opportunistic behaviors. The existence of self-organization, if turns out to exist in real mixed flow of AVs and HVs, may suggest the possibility of regulating their interactions through a decentralized approach.
The remaining part of the paper is organized as follows. We first introduce how we capture the different car-following and gap seeking behaviors of AVs and HVs through a Cellular Automata (CA) model. Then we describe the setup of simulation experiments along with results and discussion. In this section, we focus on the flux of mixed flow as well as the formation process of AV clusters in different behavioral scenarios. Self-organized behaviors are observed in both models of AV behavior. We summarize the findings and remark on future works in the last section.
Here we present a simple stochastic model to depict the essential idea behind our behavior design. We consider two classes of agents, respectively autonomous and human agents (denoted as $A$ and $H$ respectively). The agents are placed over grids a $M \times N$ grid, and each agent occupies one node.
We consider the following process. In each time step, two cells are randomly picked, and this âŠ
We are interested in the equilibrium states of this system when $t \rightarrow \infty$.
Cellular Automata (CA) is a simple yet effective framework to model agent interactions over cells, which have found wide and successful applications modeling traffic flow. In CA, each agent occupies a cell, and its behavior is determined by its own state and the state of other agents in its neighborhood. Simple functions can be applied to these cell objects to change its state and as a result it can change the states of neighboring cells
CA is a discrete computational model in traffic flow literature that comprises a system of vehicles which evolve over linear time in accordance to rules (see e.g.
In this paper, we adapt the Nagel-Schreckenberg Cellular Automaton model to introduce a model of mixed traffic flow of HVs and AVs that captures three potential behaviors of AVs - opportunistic, neighbor awareness and baseline behavior. We distinguish between the two class of vehicles in our simulation by assigning different behavioural parameters - braking probability, lane changing probability, maximum speed - and methods - traversal velocity functions - to each vehicle type and model.
To avoid confusion with the various symbols used, we provide Table
Symbol | Description |
---|---|
$p_l(X)$ | Probability of lane change for vehicle type $X$ |
$p_s(X)$ | Probability of random slowdown for vehicle type $X$ |
$v^{av}(t)$ | Velocity of AV at time-step $t$ |
$v^{av}_{max}$ | Maximum speed an AV can attain |
$v_{sl}$ | Speed limit of the road |
$s_{lv}$ | Distance to the leading vehicle |
$v^{hv}(t)$ | Velocity of HV at time-step $t$ |
$v^{hv}_{max}$ | Maximum speed an HV can attain |
$v_{aa}$ | Maximum speed of AV trailing another AV |
$v_{ah}$ | Maximum speed of AV trailing an HV |
$v_h$ | Maximum speed of HV |
$v^L$ | Velocity functions based on leading vehicle type |
$v^L_{max}$ | Maximum speed based on leading vehicle type |
$\delta(p_s)$ | Delta function (0 or 1) based on braking probability |
$\delta(p_s(AV))$ | Delta function for AV braking probability |
$P(AV)$ | Proportion of AVs in traffic |
$P(AV-AV)$ | Probability of an AV trailing another AV |
Autonomous Vehicles rely on algorithmic control for decision-making
\begin{equation} p_{l}(HV) < p_{l}(AV) \end{equation} \begin{equation} p_{s}(HV) > p_{s}(AV) \end{equation}
In our model, $p_{l}(HV)$ and $p_{s}(HV)$ create realistic, stochastic HV behavior, while $p_{l}(AV) = 1$ and $p_{s}(AV) = 0$ for AVs reflect pure opportunistic behavior. The traversal velocity update equation capturing distinct AV behavior is:
\begin{equation} v^{av}(t+1) = \min(v^{av}(t) + 1, s_{lv}, v_{max}^{av}, v_{sl}) \end{equation}
Here, $v^{av}(t) + 1$ represents acceleration.
AVs communicate through V2X technology
\begin{equation} v_{max}= \left{ \begin{array}{ll} v_{aa} & \text{if } AV-AV
v_{ah} & \text{if } AV-HV
v_{h} & \text{if } HV
\end{array} \right. \label{eq:v} \end{equation}
In this model, the following condition holds:
\begin{equation} \label{eq:cond} v_{aa} > v_{ah} \geq v_{h} \end{equation}
From Equations \ref{eq:v} and \ref{eq:cond}, an AV trailing another AV attains the highest possible speed, $v_{aa}$, while an HVâs speed is capped at $v_{h}$, independent of the leading vehicleâs class. Velocity functions $v^L$ depend on the leading vehicle type, as shown in Fig
\begin{equation} v^{av}(t+1) = \min(v^{av}(t)+1, s_{lv}, v_{max}^{L}, v_{sl}) - \delta(p_s(AV)) \end{equation}
Here, $v^{av}(t) + 1$ represents acceleration, and $\delta(p_s(AV))$ is a delta function based on AV braking probability:
\begin{equation} \label{deltas} & \delta(p_s(AV)) =
\left{ \begin{array}{ll} 1 & \text{if } p_s(AV) < X \text{ for } X \sim U(0,1)
0 & \text{otherwise}
\end{array} \right. \end{equation}
In this model, AVs behave exactly like HVs, with identical behavior between the two. Traditional CA rules from Fig
\begin{equation} \label{eq:b1} p_{l}(HV) = p_{l}(AV) \end{equation} \begin{equation} \label{eq:b2} p_{s}(HV) = p_{s}(AV) \end{equation}
Both vehicle classes have the same braking and lane-changing probabilities (Eqns. \ref{eq:b1} and \ref{eq:b2}) and follow the same traversal velocity update equation:
\begin{equation} \label{eq:bv} v(t+1) = \min(v(t) + 1, s_{lv}, v_{max}, v_{sl}) - \delta(p_s) \end{equation}
Our simulation and analysis program was written in Python3, with graphics created using the Pygame 1.9.6 module. The model was implemented using Object-Oriented Programming (OOP) principles, with the main objects being Road
and Car
. NaSch CA algorithms were implemented by passing functions to these data structures and controlling program flow. The code repository for this program is available at
After observing several simulation cases, we discovered that AVs in certain models and scenarios displayed strong self-organization phenomena. Starting from random initial states, AVs evolved into configurations where they moved close to one another for extended periods. The two most prominent types of self-organized events observed were: AVs moving closely together in different lanes and AVs moving closely together in the same lane. For these simulations, we define a cluster as follows:
\[\textit{A cluster is a group of 4 or more AVs, each at most 3 cells apart.}\]This broad definition allows us to incorporate both self-organized platooning and lane-formation events under a single collective phenomenon. The prominence of one form of clustering over the other depends on factors such as the radius of the circular road, the proportion of AVs, HV distribution, and overall traffic density. For instance, longer circular roads, higher AV proportions, localized HV distributions, and lower traffic densities increase the likelihood of lane-formation clusters compared to platooning clusters.
The chosen cluster definition (4 AVs within 3 cells) is arbitrary, suited to a simulated road of 100 cells. This captures both the proximity and visual significance of clustering. Prior works (e.g.,
We hypothesize that clusters naturally form due to the opportunistic nature of AVs, modeled through gap-seeking and braking behaviors (Section
In this section, we discuss two experiments that examine the impact of three AV modelsâneighbor aware, opportunistic, and baselineâon traffic flow and emergent behavior patterns. We further explore combinations of these models: neighbor aware and opportunistic and baseline headway (allowing AVs to maintain smaller headways than HVs).
\begin{equation}\label{eqn:awareoppo} v^{av}(t+1) = \min(v^{av}(t)+1, s_{lv}, v_{max}^{L}, v_{sl}) \end{equation}
Parameter | Neighbor Aware | Opportunistic | Neighbor Aware and Opportunistic | Baseline Headway | Baseline |
---|---|---|---|---|---|
$p_{l}(HV)$ | 0.6 | 0.6 | 0.6 | 0.6 | 0.6 |
$p_{l}(AV)$ | 0.6 | 1 | 1 | 0.6 | 0.6 |
$p_{s}(HV)$ | 0.4 | 0.4 | 0.4 | 0.4 | 0.4 |
$p_{s}(AV)$ | 0.4 | 0 | 0 | 0.4 | 0.4 |
$v_{aa}$ | 5 | 4 | 5 | 4 | 3 |
$v_{ah}$ | 4 | 4 | 4 | 4 | 3 |
$v_h$ | 3 | 3 | 3 | 3 | 3 |
Each experiment was run on a circular road with three lanes (100 cells each), where vehicles start from an empty state and are distributed randomly with zero initial velocity. The vehicle type is allocated stochastically, with $P(AV) = 0.3$, representing the proportion of AVs.
We simulated five AV models for 1200 time steps across three density regimes: 0.08 (low), 0.2 (critical), and 0.6 (high). Densities are normalized, with 1 representing jam density, and $P(AV)$ fixed at 0.3. Key parameters tracked included cluster numbers, sizes, survival times, and Clusterability (ratio of vehicles in clusters to total vehicles).
Results are summarized in Table
AV Model | Low Density Cluster Count | Low Density Survival Time | Critical Density Cluster Count | Critical Density Survival Time | High Density Cluster Count | High Density Survival Time |
---|---|---|---|---|---|---|
Opportunistic | 1.04 | 17.15 | 2.05 | 61.95 | 4.85 | 1200 |
Neighbor Aware | 1.07 | 7.01 | 1.64 | 31.45 | 4.81 | 1200 |
Baseline | 0.0 | 0.0 | 1.52 | 17.75 | 4.68 | 1200 |
Baseline Headway | 1.05 | 6.14 | 1.69 | 23.30 | 4.68 | 1200 |
Neighbor Aware & Opportunistic | 1.04 | 20.75 | 1.89 | 97.67 | 5.04 | 1200 |
Survival times revealed that the baseline model clusters were transient, whereas opportunistic models led to more prominent, longer-lasting clusters.
The AV lane-changing behavior showed that opportunistic AV models had the highest rates across all regimes. This tendency was even more pronounced in the critical density regime, where lane-changing peaked across models.
This experiment varied system density linearly to jam density, running 100 time steps per density increment of 0.03 (normalized scale). Figure
Our findings suggest that opportunistic AV models performed best overall, with neighbor aware and opportunistic AVs yielding the highest throughput. Clustering provided significant benefits, especially in low to critical densities, as AVs optimized traffic flow by maintaining shorter headways and opening space for HVs.
Lane Change Dynamics: The number of lane changes highlighted behavioral distinctions. Opportunistic models led to high lane-changing rates, with neighbor aware and opportunistic AVs exhibiting âintelligent herdingâ behavior, wherein AVs minimized HV lane changes through self-organized clusters.
This herding effect had safety, design, and fuel implications, as AV clusters facilitated smoother traffic by restricting HV lane changes.
The study verifies that prominent clustering is unique to systems involving opportunistic intelligent agents capable of recognition. Our findings suggest it is possible to design AV behaviors that form self-organized clusters, enhancing traffic efficiency.
Further research will explore implications for road safety, traffic design, and fuel efficiency due to clustering and herding dynamics.
In this paper, we investigated the behavior of mixed traffic flow consisting of AVs and HVs through a series of simulation experiments. Our primary goal was to explore the relationship between the opportunistic behaviors of AVs and the formation of clusters and lanes without centralized control. To achieve this, we developed a new cellular automata model that captures potential behavioral differences between AVs and HVs, particularly focusing on the gap-seeking behaviors and neighbor awareness of AVs. We conducted simulation experiments across two scenarios: fixed densities and linearly increasing densities, allowing us to compare the clustering process and traffic performance of three AV models.
A key finding of this research is the self-organized formation of AV clusters and lanes without centralized control, even in the absence of dedicated AV lanes. This phenomenon, observed consistently across simulations, suggests that clustering may be an intrinsic property of mixed HV and AV flows. We hypothesize that this self-organization results from incentives that encourage AVs to seek and align with other AVs. Opportunistic behaviors amplify this effect, as seen in our experiments. If confirmed through further simulation or real-world experiments, this hypothesis suggests that regulating mixed HV-AV flow might be achievable through decentralized incentives rather than centralized coordination. Our results also indicate a positive impact of clustering on overall traffic flow.
The findings in this paper provide initial evidence of self-organization in mixed AV-HV traffic flows. However, we acknowledge that our model is not quantitatively realistic. Since large-scale AV deployments are not yet in place, data on AV behaviors is limited, and the opportunistic behaviors we modeled are merely one plausible scenario. Future research, as more data becomes available, should aim to model AV decision-making more realistically and further validate the findings presented here.
This research was partially supported by a Seed Grant from the Hurricane Resilience Institute (HuRRI).
Figure 1: Road - Car grid in Cellular Automata
Rule A:
In our model, $v_{lâ}$ is defined as: \begin{equation} v_{lâ} = \min(v_{max}, d_{back}) \end{equation} Here, $v_{lâ}$ represents the potential speed a vehicle can attain if it switches from lane $l$ to lane $lâ$, and $d_{j}$ is the number of empty cells in front of the vehicle in lane $l$. The incentive criterion dictates that the vehicle will switch lanes only if: \begin{equation} v_{lâ} > d_{j} \end{equation}
Rule B:
\begin{equation} d_{back} > v_{prev} \end{equation} Once an incentive to switch lanes has been established, the safety criterion requires that the vehicle checks for the car behind it in the target lane $lâ$. The distance to this car, $d_{back}$, must be greater than the speed of the previous car $v_{prev}$ in the target lane to avoid any collision while switching.
Rule C:
\begin{equation} l \rightarrow lâ \end{equation} After satisfying Rules A and B, the vehicle has a probability, $p_l$, of switching lanes. This probability reflects the driverâs reluctance to switch lanes even when conditions permit, making the model stochastic and realistic. The vehicle changes lanes only if all conditions are met.
Rule 1 and Rule 2:
If $v_j < v_{max}$ and the distance to the next vehicle $d_j$ is greater than $v_j + 1$, the vehicle accelerates, and the new speed $v_j$ is: \begin{equation} v_j = \min(\min(v_j + 1, v_{max}), d_j) \end{equation} Rule 1 ensures that the vehicle accelerates linearly until reaching its maximum speed $v_{max}$, while Rule 2 ensures collision avoidance.
Rule 3:
\begin{equation} v_j = v_j - 1 \end{equation} The vehicle may slow down randomly based on a braking probability, $p_s$, that is dependent on vehicle type. The new speed $v_{j}$ is given by the above equation. This rule models erratic driver behavior in HVs and reflects the lack of such behavior in AVs, adding realism by removing total determinism from the simulation.
Rule 4:
The vehicle moves $v_j$ cells from its initial position in the upstream direction: \begin{equation} x_j = x_j + v_j \end{equation}
\label{pavav}
The following algorithm calculates $P(\text{AV-AV})$, which represents the probability of an AV trailing another AV. This value is calculated as the ratio of the running total of AV-AV pairs to the running total of all car pairs.
# P(AV-AV) Calculation Algorithm
def calculate_p_av_av():
update = 0
trailing_av = 0
trailing_car = 1 # Assumes circular road and num(Car) > 1
while simulation_is_running():
timestep += 1
if av_trailing_av():
trailing_av += 1
trailing_car += 1
else:
trailing_car += 1
p_av_av = trailing_av / trailing_car # Update at each timestep
return p_av_av
The GitHub repository NaSch_CA_Traffic_Flow_Analysis_Software contains all simulation and analysis software developed for this research. It also includes the data files used in this paper. ```