Abstract—This paper tries to make self-driving vehicles have
human drivers’ common sense and intuitive decision-making
ability. Human drivers often make decisions according to not
only what they see, but also their predictions based on
experiences and reasoning results. We propose a systematical
intuitive decision-making for self-driving vehicles. The method
combines similarity matching, online learning mechanism and
prediction together. Similarity matching can make a decision
based on previous learned knowledge, while online learning can
enrich the knowledge database, and prediction can make the
system have reasoning common sense to produce decisions in
unfamiliar and incomplete traffic scenarios. Basically, intuitive
decision-making can produce a decision quickly without
long-time reasoning computation. A simple test example tested
the proposed method.
I. INTRODUCTION
Human drivers’ decision-making behavior is usually
affected by many factors such as driver’s experience and skills,
vehicle states, road environment. Because of the limitation of
information processing capabilities, in some cases self-driving
vehicles cannot express and acquire knowledge from
multi-source information simultaneously
[1]
, as a result, it is
difficult for them to act as accurate and quick as human drivers.
Human drivers make driving decisions not only according to
what they see and hear, but common sense based on their
many years’ driving experience and reasoning cues.
A. Previous Related Work
There have been a lot of studies on decision making of
robotic vehicles. Sun-Do Kim et al. proposed a fuzzy-based
decision making algorithm for outdoor navigation of mobile
robots to satisfy the robot's safety requirements and to spend
less time on the intersection
[2]
. Wang proposed a fuzzy neural
network model of decision mechanism of driving system on
freeway
[3]
. Chih-Li Huo focused on vehicle warning system
based on fuzzy decision making for lane departure and
forward collision avoidance for driver assistant system. The
proposed system is composed of two parts: vision-based
preprocessing and a fuzzy decision making
[4]
.
Sandor M Veres et al. described a brief description of
Markova decision process
[5]
. Based on this work, A. Foka
developed a hierarchical formulation of a partially observable
Markov decision processes (POMDP) for autonomous robot
real-time navigation in dynamic environments
[6]
. Similarly,
Augie W adopted POMDP model as a decision making model
*Resrach supported by the National Natural Science Foundation of China
(No. 61304194 and No. 51275041).
The authors are with Beijing Institute of Technology, Beijing, China.
(gongjianwei@bit.edu.cn).
for autonomous vehicle with a predictive framework. The
belief states resulting from N-steps prediction were used to
determine the best action policy
[7]
. POMDP could address the
problem of the agent to act optimally in partially observable
environment
[6]
.
Similarly, Linjin Wu et al. put forward an autonomous
decision-making agent model for autonomous vehicles, which
mainly included: data pretreatment, decision-making response,
decision-making execution, decision-making evaluation and
learning feedback
[8]
. Also, a multiple-criteria decision-making
(MCDM) approach was presented to divide the decision
process into two stages: while the first one determined feasible
maneuvers based on Petri nets, and the second selected the
most appropriate maneuver using a multivariate utility
function
[9]
. However, the approach did not explicitly consider
environment uncertainties.
Jerzy et al. proposed two approaches based on proactive
and reactive scheduling. In the former, it was assumed that
tasks’ execution times were not known precisely, and the
minimal absolute regret decision making problem was solved.
The latter used motion control decisions for the group of
vehicles
[10]
. While Ralf Regele described how a high-level
abstract world model could be used to support the
decision-making process of an autonomous driving system.
The approach proposed a hierarchical world model which
included a low-level model for the trajectory planning and a
high-level model for solving the traffic coordination
problem
[11]
.
R. R. Yager and F. E. Petry presented an intuitive
decision- making method, using hyper similarity matching for
modeling the situational matching with importance weights
for situations with different weight vectors
[12]
. As shown in
Figure 1, the method made an intuitive decision by matching
the current problem P
n
with the solutions in the knowledge
database, and found the S
2
was the best match for P
n
.
Current problem
Pn
Previous situations
S
1, S2, …, Sm
Intuitive mathcing
Possible solution generated
S
2-Pn
Figure 1. Intuitive Decision Modeling Flow
However, the intuition decision-making model here is still
a simple model that cannot carry on machine learning,
furthermore it cannot adapt to the complex traffic environment.
Jianwei Gong* Member, IEEE , Shengyue Yuan, Jiang Yan, Xuemei Chen, Huijun Di
Intuitive Decision-making Modeling for Self-driving Vehicles
2014 IEEE 17th International Conference on
Intelligent Transportation Systems (ITSC)
October 8-11, 2014. Qingdao, China
978-1-4799-6078-1/14/$31.00 ©2014 IEEE 29