arXiv:2204.03656v2 [cs.

RO] 2 Nov 2022

Automatic Hyperparameter Optimization

Using Genetic Algorithm in Deep
Reinforcement Learning for Robotic
Manipulation Tasks
Adarsh Sehgal1,2 , Nicholas Ward1,2 , Hung La1,2* and Sushil
Louis2
1Advanced Robotics and Automation (ARA) Laboratory.
2 Department of Computer Science and Engineering, University
of Nevada, Reno, 89557, NV, USA.

*Corresponding author(s). E-mail(s): hla@unr.edu;

Contributing authors: asehgal@nevada.unr.edu;
nhward@nevada.unr.edu; sushil@cse.unr.edu;

Abstract
Learning agents can make use of Reinforcement Learning (RL) to decide
their actions by using a reward function. However, the learning pro-
cess is greatly influenced by the elect of values of the hyperparameters
used in the learning algorithm. This work proposed a Deep Determin-
istic Policy Gradient (DDPG) and Hindsight Experience Replay (HER)
based method, which makes use of the Genetic Algorithm (GA) to fine-
tune the hyperparameters’ values. This method (GA+DDPG+HER)
experimented on six robotic manipulation tasks: FetchReach; FetchSlide;
FetchPush; FetchPick&Place; DoorOpening; and AuboReach. Analysis
of these results demonstrated a significant increase in performance
and a decrease in learning time. Also, we compare and provide evi-
dence that GA+DDPG+HER is better than the existing methods.

Keywords: DRL, Reinforcement Learning, Genetic Algorithm,

GA+DDPG+HER, DDPG, HER

1
2 Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforcem

1 Introduction
Reinforcement Learning (RL) [1] has recently been applied to a variety of
applications, including robotic table tennis [2], surgical robot planning [3],
rapid motion planning in bimanual suture needle regrasping [4], and Aquatic
Navigation [5]. Each one of these applications employs RL as a motivating
alternative to automating manual labor.
Studies have shown that hyperparameter tuning, particularly when uti-
lizing machine learning, can have a significant impact on an algorithm’s
performance [6, 7]. From this came the inspiration to enhance the RL algo-
rithm. While the hill climbers [8] can be useful in specific situations, they are
not very useful in complex situations like robotic tasks employing RL. This is
due to the lack of a clear correlation between performance changes and chang-
ing the settings of hyperparameters. The number of epochs it takes the learning
agent to learn a given robotic task can be used to assess its performance. Due
to the non-linear nature of the relationship between the two, we looked into
using a Genetic Algorithm (GA) [9], which may be utilized for hyperparameter
optimization. Even while GAs can have significant computational costs, which
could be an overhead for any RL algorithm, GAs offer a potential method to
tune the hyperparameters in comparable kinds of issues once and for all. GA
has the ability to adjust the hyperparameters once and utilize them endlessly
in subsequent RL trainings, saving computational resources that would oth-
erwise be spent each time an RL is applied to an agent. [10] demonstrates
how GA is preferred over other optimization techniques. This is due to GA’s
capacity to handle either discrete or continuous variables, or both. It enables
the simultaneous evaluation of n members of a population, enabling multipro-
cessor machines to run parallel simulations. GA is inherently suitable for the
solution of multi-objective optimization problems since it works with a pop-
ulation of solutions [11]. [12] evaluated a small number of cost functions to
compare the performance of eight algorithms in solving simple and complex
building models. They discovered that the GA frequently approached the opti-
mal minimum. This demonstrates that GA is a powerful algorithm for solving
complex problems, making it our top pick for hyperparameter optimization.
[13], [14], [15], and [16] are some of the closely related works. These
publications’ findings add to the growing body of data that using a GA to auto-
matically modify the hyperparameters for DDPG+HER can greatly enhance
efficiency. The discrepancy can have a notable impact on how long it takes a
learning agent to learn.
We develop a novel automatic hyperparameter tweaking approach in this
paper, which we apply to DDPG+HER from [17]. The algorithm is then
applied to four existing robotic manipulator gym environments as well as three
custom-built robotic manipulator gym environments. Furthermore, the entire
algorithm is examined at various stages to determine whether the technique is
effective in increasing the overall efficiency of the learning process. The final
results support our claim and provide sufficient evidence that automating the
hyperparameter tuning procedure is critical since it reduces learning time by
Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforcement L

up to 57%. Finally, we compare GA+DDPG+HER to four different approaches

used on FetchReach. GA+DDPG+HER outperforms them all. Open source
code is available at https://github.com/aralab-unr/ga-drl-aubo-ara-lab.
The following is a list of our major contributions:
• A novel algorithm for automatic hyperparameter tweaking.
• Six simulated and one real task were used to test the algorithm. To
analyze the algorithm, we created Aubo-i5 simulated and actual custom
environments.
• As the GA develops, the training process is evaluated based on a variety of
hyperparameters.
• The efficiency of DDPG+HER was tested over ten runs using
GA+DDPG+HER discovered hyperparameters in both simulated and
actual manipulation tasks.
• GA+DDPG+HER was compared to other approaches.

2 Background
2.1 Genetic Algorithm (GA)
Genetic Algorithms (GAs) [9, 18, 19] were created to explore poorly-
understood areas [20], where an exhaustive search is impossible and other
search methods perform poorly. GAs, when employed as function optimizers,
aim to maximize fitness that is linked to the optimization goal. On a range of
tough design and optimization issues, evolutionary computing techniques in
general, and GAs in particular, have had a lot of empirical success. They begin
with a population of candidate solutions that have been randomly initialized
and are commonly encoded in a string (chromosome). A selection operator
narrows the search space to the most promising locations, whereas crossover
and mutation operators provide new potential solutions.
To choose parents for crossover and mutation, we employed ranking
selection [21]. Higher-ranked (fitter) individuals are probabilistically selected
through rank selection. Unlike fitness proportionate selection, ranking selection
is concerned with the existence of a fitness difference rather than its magni-
tude. Uniform crossover [22] is used to create children, who are then altered
by flip mutation [19]. Binary coding with concatenated hyperparameters is
used to encode chromosomes. [23] shows one such example of GA paired with
Lidar-monocular visual odometry (LIMO).

2.2 Deep Reinforcement Learning (DRL)

Q-learning [24] approaches have been used by autonomous robots to do a
range of tasks [25], and significant study has been done in this field since its
inception [24], with some work focused on continuous action spaces [26–29]
and others on discrete action spaces [30]. Reinforcement Learning (RL) [1] has
been used to improve locomotion [31, 32] and manipulation [33, 34]. In the field
of robotic research, deep reinforcement learning (DRL) has become a potent
4 Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforcem

control method [35]. The DRL is more strong in its detailed examination of the
environment than in control theory. When used on robots, this DRL capacity
produces more intelligent and human-like behavior. Robots can thoroughly
examine the environment and discover useful solutions when DRL techniques
are used in conjunction with adequate training [36].
There are two types of RL methods: off-policy and on-policy. On-policy
methods seek to assess or improve the policy upon which decisions are made,
for e.g. SARSA learning [37]. For real robot systems, off-policy methods [38]
such as the Deep Deterministic Policy Gradient algorithm (DDPG) [39], Prox-
imal Policy Optimization [40], Advantage Actor-Critic (A2C) [41], Normalized
Advantage Function algorithm (NAF) [42] are useful. There is also a lot of work
on robotic manipulators [43, 44]. Some of this work relied on fuzzy wavelet
networks [45], while others relied on neural networks [46, 47] to complete their
goals. [48] provides a comprehensive overview of modern deep reinforcement
learning (DRL) algorithms for robot handling. Goal-conditioned reinforcement
learning (RL) frames each activity in terms of the intended outcome and, in
theory, has the potential to teach a variety of abilities [49]. The robustness
and sample effectiveness of goal-reaching procedures are frequently enhanced
by the application of hindsight experience replay (HER) [50]. For our trials,
we are employing DDPG in conjunction with HER. [51] describes recent work
on applying experience ranking to increase the learning pace of DDPG+HER.
Both a single robot [52, 53] and a multi-robot system [54–58] have been
extensively trained/taught using RL. Both model-based and model-free learn-
ing algorithms have been studied previously. Model-based learning algorithms
are heavily reliant on a model-based teacher to train deep network policies in
real-world circumstances.
Similarly, there has been a lot of work on GA’s [9] [59] and the GA operators
of crossover and mutation [60], which have been applied to a broad variety of
problems. GA has been used to solve a wide range of RL problems [60–63].

2.3 Proximal Policy Optimization (PPO)

Policy Gradient (PG) [64] approaches are one of several RL techniques used
to optimize the policy. These methods employ the rewards to compute an
estimator of the policy gradient and plug it into a stochastic gradient ascent
algorithm [65] . The Proximal Policy Optimization (PPO) [40] technique can
be applied to environments with either continuous or discrete action spaces
[66]. For reinforcement learning, PPO is a family of Policy Gradient algorithms
[67]. Without any prior knowledge, it is a model-free reinforcement learning. A
DRL agent is said to be model-free if it hasn’t learned an explicit model of its
surroundings. A surrogate objective for computing policy updates is the main
component of the PPO algorithm. In the spirit of a trust region strategy, the
surrogate objective regulates significant policy updates so that each step stays
within a near vicinity of the previous-iteration policy [68].
Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforcement L

2.4 Advantage Actor Critic (A2C)

Advantage actor critic, often known as A2C, is a synchronous variant of Asyn-
chronous advantage actor-critic (A3C) models that performs as well as or
better than the asynchronous version [40]. A2C [41] technique is based on pol-
icy gradients. By forecasting what actions are good in advance, it dissociates
value estimates from action selection. Only effective revisions to the policy are
produced using value estimation. In essence, it immediately raises the likeli-
hood of positive behaviors and lowers the likelihood of negative ones [69]. [70]
The policy estimation and value function estimate procedures used in actor-
critic based DRL models apply to an advantage function. The A2C approach
estimates the policy function with a critic network and the Q-value function
with an actor network instead of just using a single learner neural network to
estimate the Q-value function. Actor and critic networks are updated simulta-
neously using A2C actor-critic models. The calculation time is longer for A3C
models since they estimate both actor and critic networks simultaneously and
asynchronously. We would be using A2C in our experiments because there is
not much of a performance difference between synchronous and asynchronous
actor-critic models.

2.5 GA on DDPG+HER
GA can be used to solve a variety of optimization problems as a function
optimizer. This study concentrates on the DDPG+HER, which was briefly
discussed earlier in this chapter. Based on their fitness values, GAs can be
utilized to optimize the hyperparameters in the system. GA tries to achieve the
highest level of fitness. Various mathematical formulas can be used to convert
an objective function to a fitness function.
Existing DDPG+HER algorithms have a set of hyperparameters that can-
not be changed. When GA is applied to DDPG+HER, it discovers a better set
of hyperparameters, allowing the learning agent to learn more quickly. The fit-
ness value for this problem is the inverse of the number of epochs. GA appears
to be a potential technique for improving the system’s efficiency.

3 Genetic Algorithm optimization for Deep

Reinforcement Learning
3.1 Reinforcement Learning
Consider a typical RL system, which consists of a learning agent interacting
with the environment. [50] S is the set of states, A is the set of actions, p(s0 ) is
a distribution of initial states, r : S ×A →
− R is a reward function, p(st+1 |st , at )
are transition probabilities, and γ ∈ [0, 1] is a discount factor that can be used
to represent an environment.
A deterministic policy depicts the relationship between states and actions
as follows: π : S →− A. Every episode begins with the sampling of the initial
6 Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforcem

state s0 . The agent acts at for each timestep t based on the current state st:
st : at = π(st ). The accomplished action is rewarded with rt = r(st , at ), and
the distribution
P∞ p(.|st , at ) aids in sampling the new state of the environment.
Rt = i=T γ i−t ri is the discounted sum of future rewards. The purpose of the
agent is to maximize its expected return E[Rt |st , at ], and an optimal policy
∗
can be defined as any policy π ∗ , such that Qπ (s, a) ≥ Qπ (s, a) for each s ∈
S, a ∈ A, and any policy π. An optimal Q-function, Q∗ , is a policy that has
the same Q-function as the best policy and fulfills the Bellman equation:

Q∗ (s, a) = Es0 p(.|s,a)) [r(s, a) + γmax

0
Q∗ (s0 , a0 ))]. (1)
a ∈A

3.2 Deep Deterministic Policy Gradients (DDPG)

There are two neural networks in Deep Deterministic Policy Gradients [39]
(DDPG): an Actor and a Critic. The critic neural network is an action-value
function approximator Q : S × A → − R and the actor neural network is a target
policy π : S →− A. With weights θQ and θµ , the critic network Q(s, a|θQ ) and
actor network µ(s|θµ ) are randomly initialized.
By offering a reliable target to the training process, the target network,
which is typically created as a replica of the critic network, is also used for
the purpose of updating the critic network. To produce episodes, a behavioral
policy is used, which is a noisy variation of the target policy, πb (s) = π(s) +
N (0, 1). A critic neural network is trained similarly to a DQN, except that the
target yt is computed as yt = rt +γQ(st+1 , π(st+1 )), where γ is the discounting
factor. The actor network is trained using the loss La = −Ea Q(s, π(s)). The
experience replay and target network are both critical for sustaining the DDPG
method’s training and enabling deep neural networks.

3.3 Hindsight Experience Replay (HER)

Hindsight Experience Reply (HER) [50] aims to learn from setbacks by imi-
tating human behavior. Even if the agent does not achieve the intended goal,
the agent learns from each experience. HER considers the modified goal to be
whatever condition the agent reaches. Only the transition (st g, at , rt , st+1 g)
with the original aim g is saved in standard experience replay. The transition
(st g 0 , at , rt0 , st+1 g 0 ) to modified goal g 0 is also stored by HER. HER performs
admirably with extremely scarce incentives and is far superior to shaped
rewards in this regard.

3.4 Open Problem Discussion

The efficiency of DDPG + HER is a concern. Using a better set of hyper-
parameters in the algorithm can increase the performance of the majority of
robotic activities. The number of epochs for the learning agent to learn spe-
cific robotic tasks can be used to assess performance. The remaining sections
Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforcement L

of this section demonstrate how changing the values of various hyperparame-

ters has a major impact on the agent’s competency. Later in this section, the
solution to this problem is described, and the accompanying experimental find-
ings show that the proposed approach outperforms the existing reinforcement
learning technique.

3.5 DDPG + HER and GA

The main contribution of our paper is presented in this section: the genetic
algorithm explores the space of hyperparameter values utilized in DDPG +
HER for values that maximize task performance while reducing the number
of training epochs. The following hyperparameters are targeted: discounting
factor γ; polyak-averaging coefficient τ [71]; learning rate for critic network
αcritic ; learning rate for actor-network αactor ; percent of times a random action
is taken ; and standard deviation of Gaussian noise added to not completely
random actions as a percentage of the maximum absolute value of actions on
different coordinates η. All of the hyperparameters have a range of 0-1, which
may be justified using the equations in this section.
Adjusting the values of hyperparameters did neither boost nor reduce the
agent’s learning in a linear or immediately visible pattern, according to our
experiments. As a result, a basic hill climber is unlikely to identify optimal
hyperparameters. We use our GA to optimize these hyperparameter values
because GAs were created for such poorly understood problems.
We use τ , the polyak-averaging coefficient, to show the performance non-
linearity for different values of τ . As seen in Equation (2), τ is employed in the
algorithm:

0 0
θQ ←
− τ θQ + (1 − τ )θQ ,
0 0
θµ ←
− τ θµ + (1 − τ )θµ . (2)

Equation (3) demonstrates how γ is employed in the DDPG + HER

method, whereas Equation (4) explains the Q-Learning update. α denotes the
rate at which you are learning. This update equation is used to train networks.

0 0
yi = ri + γQ0 (si+1 , µ0 (st+1 |θµ )|θQ ), (3)

Q(st , at ) ←
− Q(st , at ) + α[rt+1 + γQ(st+1 , at+1 )
−Q(st , at )]. (4)

We will need two learning rates, one for the actor-network (αactor ) and the
other for the critic network (αcritic ) because we have two types of networks.
8 Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforcem

The use of the percent of times that a random action is executed  is explained
by equation (5).

(
a∗t with probability 1 − ,
at = (5)
random action with probability .

Figure 1 indicates that changing the value of τ causes a change in the

agent’s learning, underscoring the importance of using a GA. The comparison
of 11 τ values is shown in the first plot. For clarity, two additional charts com-
pare three τ values. We are using four CPUs and the original (untuned) value
of τ in DDPG was 0.95. All values are taken into account up to two decimal
places to examine how the success rate changes as the hyperparameter values
change. We can see from the plots that there is a lot of room for improvement
from the original success rate.
The integration of DDPG + HER with a GA, which uses a population
size of 30 across 30 generations, is explained in Algorithm 1. The number of
experiments with various values was used to determine population size and
the number of generations. These values are sufficient to compute the success
rate for a sufficient number of chromosomes. To choose parents, we used rank-
ing selection [21]. The parents are determined probabilistically based on rank,
which is determined by relative fitness (performance). Then, using uniform
crossover [22], children are created. We are also utilizing flip mutation [19]
with a 0.1 mutation frequency. Each hyperparameter is encoded using a binary
chromosome, and the bits are concatenated to generate a chromosome for the
GA. Polyak-averaging coefficient, discounting factor, learning rate for critic
network, learning rate for actor-network, percent of times a random action
is taken, and standard deviation of Gaussian noise added to not completely
random actions as a percentage of the maximum absolute value of actions
on different coordinates are the six hyperparameters in order. We need 66
bits for six hyperparameters, since each hyperparameter requires 11 bits to be
expressed to three decimal places. Domain-independent crossover and muta-
tion string operators can then generate new hyperparameter values using these
string chromosomes. Because tiny changes in hyperparameter values produce
significant variations in success rates, we examined hyperparameter values up
to three decimal places. A step size of 0.001 is, for example, considered the
greatest fit for our problem. It might be important to mention that every
chromosome was examined exactly once. Even if the results of identical chro-
mosomal runs may vary slightly, we think it is safe to assume that the fitness
function has a low variance, which keeps the calculation cost low.
The inverse of the number of epochs it takes for the learning agent to
achieve close to maximum success rate (≥ 0.85) for the first time determines
the fitness of each chromosome (set of hyperparameter values). Because GA
always maximizes the objective function, fitness is inversely proportional to the
number of epochs, converting our minimization problem into a maximization
Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforcement L

Fig. 1: Success rate vs. epochs for various τ for FetchPick&Place-v1 task.
10 Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforce

Algorithm 1 Proposed GA+DDPG+HER Algorithm

1: Choose a population of n chromosomes
2: Set the values of hyperparameters into the chromosome
3: Run the DDPG + HER to get the number of epochs for which the
algorithm first reaches success rate ≥ 0.85
4: for all chromosome values do
5: Initialize DDPG
6: Initialize replay buffer R ← φ
7: for episode=1, M do
8: Sample a goal g and initial state s0
9: for t=0, T-1 do
10: Sample an action at using DDPG behavioral policy
11: Execute the action at and observe a new state st+1
12: end for
13: for t=0, T-1 do
14: rt := r(st , at , g)
15: Store the transition (st g, at , rt , st+1 g) in R
16: Sample a set of additional goals for replay G := S(current
episode)
17: for g 0 ∈ G do
18: r0 := r(st , at , g 0 )
19: Store the transition (st g 0 , at , r0 , st+1 g 0 ) in R
20: end for
21: end for
22: for t=1, N do
23: Sample a minibatch B from the replay buffer R
24: Perform one step of optimization using A and minibatch B
25: end for
26: end for
27: return 1/epochs
28: end for
29: Perform Uniform Crossover
30: Perform Flip Mutation at a rate of 0.1
31: Repeat for the required number of generations to find an optimal solution

problem. We utilize a GA search since an exhaustive search of the 266 -size

search area is not practicable due to the length of each fitness evaluation.

4 Experimental Results
4.1 Experimental setup
A chromosome is binary encoded, as previously stated. Each chromosomal
string is the result of combining all of the GA’s arguments. Figure 4 depicts
an example chromosome with four binary encoded hyperparameters.
Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforcement L

(a) GA+DDPG+HER over ten runs, vs. Original

(b) GA+DDPG+HER averaged over ten runs, vs. Original

Fig. 2: Success rate vs. epochs for FetchPush-v1 task when τ and γ are found
using the GA.

FetchPick&Place-v1, FetchPush-v1, FetchReach-v1, FetchSlide-v1, and

DoorOpening are the environments used to test robot learning on five different
simulations tasks (see Figures 5 and 6). Figures 7 and 8 depict AuboRe-
ach habitats, which are used in both simulated and real-world research. On
these six gym situations, we test our algorithm. FetchPick&Place, FetchPush,
FetchReach, and FetchSlide are fetch environments from [72]. We created
DoorOpening and AuboReach, which are custom-built gym environments. The
following are the details of the six tasks:
• FetchPick&Place: The agent picks up the box from a table and moves to
the goal position, which may be anywhere on the table or the area above it.
• FetchPush: In front of the agent, a box is stored. It pushes or rolls the box
to the table’s goal location. The agent is not permitted to take possession
of the box.
• FetchReach: In the area around the end-effector, the agent must move it
to the goal position.
• FetchSlide: The puck is placed on a slippery surface within reach of the
agent. It must strike the puck with sufficient force that it (the puck) comes
to a halt in front of the goal owing to friction.
12 Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforce

(a) GA+DDPG+HER over 2 runs, vs. Original

(b) GA+DDPG+HER averaged over 2 runs, vs. Original

Fig. 3: Success rate vs. epochs for FetchSlide-v1 task when τ and γ are found
using the GA.

Fig. 4: Chromosome representation for the GA.

Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforcement L

(a) FetchPick&Place environment (d) FetchPick&Place plot

(e) FetchPush plot

(b) FetchPush environment

(f) FetchReach plot

(c) FetchReach environment

Fig. 5: When all six hyperparameters are identified by GA, the matching
DDPG+HER versus GA+DDPG+HER charts are generated. Over ten runs,
all graphs are averaged.

• DoorOpening: A simulated Aubo i5 manipulator is positioned within reach

of a door, with the door handle pointing in the direction of the robot. The
goal is to force the door open by applying force to the area of the door
handle.
14 Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforce

(a) Door Opening environment (c) DoorOpening plot

(d) FetchSlide plot

(b) FetchSlide environment

Fig. 6: When all six hyperparameters are found by GA, the environments and
related DDPG+HER versus GA+DDPG+HER charts are shown. Over ten
runs, the graphs are averaged.

• AuboReach: An Aubo i5 manipulator, simulated or real, learns to achieve

a desired joint configuration and pick up an object with a gripper.

4.2 Running GA
We ran the GA independently on each of these scenarios to test the effec-
tiveness of our method and compared results to the hyperparameters’ original
values. We would anticipate that when comparing GA+DDPG+HER with
DDPG+HER, PPO, A2C, and DDPG, the algorithm with the fewest episodes,
running time, steps, and average epochs will do the best. This would illustrate
why optimizing hyperparameters is crucial as opposed to using the algo-
rithms’ built-in default hyperparameters. Figure 2a depicts the outcome of our
FetchPush-v1 experiment. We used the GA to run the system and discover
the best values for the hyperparameters τ and γ. We display results from ten
GA runs because the GA is probabilistic, and the results show that optimized
hyperparameters determined by the GA can lead to greater performance. The
Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforcement L

Fig. 7: Using the most accurate policy learned via GA+DDPG+HER, the
AuboReach environment performs a task in a real experiment.

Fig. 8: In a simulated experiment, the AuboReach environment performs a

task using the best policy learned via GA+DDPG+HER.

learning agent can learn faster and achieve a higher success rate. Figure 2b
depicts one learning run for the initial hyperparameter set, as well as the aver-
age learning for the ten GA iterations. As we evaluated the genetic algorithm,
the results displayed in figure 2 demonstrate changes when only two hyperpa-
rameters are tuned. We can see the potential for performance improvement.
16 Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforce

Our findings from optimizing all five hyperparameters back up our optimism,
and they are detailed below.
Figure 3 (b) shows a comparison of one original experiment with two aver-
aged runs for optimizing hyperparameters τ and γ. Because this operation can
take several hours to complete in a single run, and because it was a compo-
nent of one of our initial tests, we only performed it twice. As we evaluated the
genetic algorithm, the results displayed in Figures 2 and 3 demonstrate changes
when only two hyperparameters are tuned. We can see the potential for per-
formance improvement. Our findings from optimizing all five hyperparameters
back up our optimism, and they are detailed below.
After that, GA was used to optimize all hyperparameters, with the results
presented in Figures 5 and 6 for each task. Table 1 compares the GA-discovered
hyperparameters to the RL algorithm’s initial hyperparameters. All the sim-
ulated environments, with the exception of AuboReach, used the same set
of hyperparameters discovered by GA+DDPG+HER. As seen in table 1, yet
another set of hyperparameters was generated for AuboReach. The learning
rates, αactor and αcritic are the same as they were in the beginning, while the
other four hyperparameters have different values. Figures 5 and 6 indicate that
the GA-discovered hyperparameters outperformed the original hyperparame-
ters, suggesting that the learning agent was able to learn more quickly. All of
the plots in the previous figure have been averaged over ten runs.

All envi- Aubo-i5 - Aubo-i5 -

ronments Fixed Ini- Random
except tial and Initial and
Aubo-i5 Target Target
state state
hyper- DDPG+ GA+ GA+ GA+
parameters HER DDPG+ DDPG+ DDPG+
HER HER HER
γ 0.98 0.928 0.949 0.988
τ 0.95 0.484 0.924 0.924
αactor 0.001 0.001 0.001 0.001
αcritic 0.001 0.001 0.001 0.001
 0.3 0.1 0.584 0.912
η 0.2 0.597 0.232 0.748
Table 1: DDPG+HER vs. GA+DDPG+HER values of hyperparameters.

Figures 7 and 8 show how GA+DDPG+HER settings (as listed in table

1) were applied to a custom-built gym environment for the Aubo-i5 robotic
manipulator. The motors in this environment are controlled by the MOVEit
package [73] , but the DDPG+HER works as a brain for movement. Each
action (series of joint states) that the robot should take is determined by
DDPG+HER. The results were at first unexpected. Each epoch took several
hours to complete (> 10-15 hours). We did not complete the entire curricu-
lum because it could take several weeks. The same holds for DDPG+HER
Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforcement L

settings. This is because the movement speed of the Aubo i5 robotic manipu-
lator was kept sluggish in both simulation and real-world studies to avoid any
unexpected sudden movements, which could result in harm. In the AuboRe-
ach setting, there were also planning and execution processes involved in the
successful completion of each action. AuboReach, unlike the other gym environ-
ments covered in this study, could only run on a single CPU. This is because
other environments were introduced in MuJoCo and could easily run with the
maximum number of CPUs available. MuJoCo can create several instances
for training, which allows for faster learning. AuboReach must only perform
one action at a time, much like a real robot. Because of these characteristics,
training in this setting takes a long period.

4.3 Modifications required for AuboReach

The GA+DDPG+HER settings were then applied to the AuboReach envi-
ronment, but only to action values. This means that in both simulated and
real-world studies, the robot did not run to do the action. This further means
that whatever decision the DDPG+HER makes, the robot is taken to have
made the decision. This is further supported by the fact that MOVEit handles
the hassle of communicating move signals to numerous joints. This is accurate
since the robot can do each action determined by the DDPG+HER algorithm.
This is due to the planning and execution required for internal joint movement
in the robot [74]. This also eliminates the possibility of a collision, which may
have occurred if these procedures had been skipped. Each epoch now took
less than a minute to complete, a considerable reduction in training time that
allows this environment to be used for training.
The GA+DDPG+HER settings (table 1) were re-applied now that some
of the environmental challenges had been overcome. These hyperparameters
did not outperform the original hyperparameters in any way. We believe this
is due to the fact that this environment is far more intricate and unique than
others. We took into account even more aspects to ensure that the environment
is trainable. For training and testing, this environment employs four joints
(instead of six). shoulder, forearm, upper-arm, and wrist1 are the ones used.
This was done to ensure that the learning could be finished in a reasonable
amount of time. Each joint has a range of -1.7 to 1.7 radians. The robot’s
initial and reset states were both set to upright, i.e., [0, 0, 0, 0].
The GA+DDPG+HER method was tweaked for better learning and faster
hyperparameter search, in addition to minor adjustments to the environment.
The definition of success was changed to include ten successful epochs. This
means that the GA was regarded successful if it had a 100% success rate for
ten consecutive epochs. When αactor and αcritic are both greater than 0.001,
learning never converges, according to experiments. As a result, the values of
αactor and αcritic were set to 0.001. Multi-threading can occur when only action
values are employed; hence four CPUs could be utilized. If the cumulative
discrepancy between the target and the achieved joint states is less than 0.1
radians, AuboReach considers the DDPG+HER-determined joint states to be
18 Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforce

(a) Trained initial and target state

(b) Training is done with random initial and target joint states with 1 CPU

Fig. 9: The AuboReach task’s success rate against epochs. This graph
represents the average of over ten runs.

a success. [-0.503, 0.605, -1.676, 1.391] were chosen as the objective joint states.
We were able to find a new set of hyperparameters using these changes to
the algorithm, as shown in Table 1. Figure 9a shows the difference in success
rates between DDPG+HER and GA+DDPG+HER during training. Without
a doubt, the GA+DDPG+HER outperforms the DDPG+HER.
After the GA+DDPG+HER had determined the best hyperparameters for
the AuboReach environment, the training was repeated using four CPUs to
determine the best policy. The robots were then subjected to this policy in
both simulated and real-world testing. The crucial point to remember here
is that for testing purposes, CPU consumption was reduced to one. In both
studies, the robot was able to transition from the training’s beginning joint
space to the goal joint space. Because unpredictability was not added during
training, the environment was constrained to only one possible path. Since
both DDPG+HER and GA+DDPG+HER eventually achieved a 100% suc-
cess rate, there was no discernible difference throughout testing. The main
distinction is the speed with which the environment may learn given a set of
hyperparameters.
The AuboReach environment was updated in another experiment to train
on random joint states. The robot may now start and reach objectives in
Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforcement L

various joint states during testing thanks to the update. The GA was run on
this environment, and the hyperparameters discovered by the GA are shown in
table 1. Figure 9b shows that the plot of GA+DDPG+HER is still better than
DDPG+HER. Figures 7 and 8 show the robot in action as it accomplishes the
task of picking the object in real and simulated tests.
The use of GA+DDPG+HER, in the AuboReach environment resulted
in automatic DDPG+HER hyperparameter adjustment, which improved the
algorithm’s performance.

4.4 Training evaluation

Monitoring the status of a GA while it is operating is critical for optimizing
the system’s performance. Because of the way GAs work, certain chromosomes
will perform better than others. As a result, the performance graph is unlikely
to be a smooth rise curve. Some fitness function assessments result in a zero,
implying that the chromosome is unsuited for use. Despite the non-smooth
slope, it is believed that the system’s overall performance would improve as
GA advances.
We created numerous charts to track GA’s progress while we looked for the
best hyperparameters. Figures 10, 11, and 12 show how the system’s perfor-
mance improves as GA progresses. Median success rate across fitness function
evaluations, total reward across episodes, and epochs to accomplish the tar-
get throughout fitness function evaluations are all hyperparameters taken into
account when evaluating training performance. It can be seen that the sys-
tem’s overall performance is improving. With fitness function evaluations, the
overall reward increases as the agent’s episodes and epochs to accomplish the
goal decrease. This indicates that the GA is on the right road in terms of deter-
mining the best hyperparameter values. We plotted data for only one GA run
and limited the duration of a GA run because each GA run requires several
hours to many days of run time. When we started to see results, we decided
to terminate the GA.
Now that the GA has behaved as intended, we will evaluate the system’s
efficiency using the GA’s discovered hyperparameters.

4.5 Efficiency evaluation

We generated data for various hyperparameters to evaluate and compare the
efficiency of the GA+DDPG+HER algorithm for training the agent to accom-
plish a task. These hyperparameters are good indicators of the algorithm’s
efficiency. For the majority of the training tasks, as shown in Figure 13, the
total reward has increased significantly. The DDPG+HER algorithm’s effi-
ciency was significantly improved when rewards were raised. Because it is
steered considerably faster towards the desired task, the agent can learn a lot
faster. To get an unbiased review, we averaged these plots over ten runs. With
GA+DDPG+HER, the FetchSlide environment performed worse. We believe
that this is due to the task’s complexity. The maximum number used for that
20 Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforce

(a) FetchPick&Place - Median reward vs (d) FetchPush - Median reward vs Times

Times GA fitness function evaluated GA fitness function evaluated

(b) FetchPick&Place - Epochs vs Times (e) FetchPush - Epochs vs Times GA

GA fitness function evaluated fitness function evaluated

(c) FetchPick&Place - Median success (f) FetchPush - Median success rate vs

rate vs Times GA fitness function evalu- Times GA fitness function evaluated
ated

Fig. 10: When GA finds all six hyperparameters, the GA+DDPG+HER

training evaluation charts. This is the outcome of a single GA run.

hyperparameter was used to represent in the tables the tasks that did not
attain the target during training.
Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforcement L

(a) FetchSlide - Median reward vs Times (d) DoorOpening - Median reward vs

GA fitness function evaluated Times GA fitness function evaluated

(b) FetchSlide - Epochs vs Times GA (e) DoorOpening - Epochs vs Times GA

fitness function evaluated fitness function evaluated

(c) FetchSlide - Median success rate vs (f) DoorOpening - Median success rate
Times GA fitness function evaluated vs Times GA fitness function evaluated

Fig. 11: When all six hyperparameters are identified by GA, the
GA+DDPG+HER training evaluation plots. One GA run yielded this result.

We also generated more data to evaluate the episodes, running time (s),
steps, and epochs that an agent must learn to achieve the desired goal. Tables
2-5 present this information. The information in the tables is an average of
22 Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforce

(a) AuboReach - Median reward vs (d) FetchReach - Median reward vs

Times GA fitness function evaluated Times GA fitness function evaluated

(b) AuboReach - Epochs vs Times GA (e) FetchReach - Epochs vs Times GA

fitness function evaluated fitness function evaluated

(c) AuboReach - Median success rate vs (f) FetchReach - Median success rate vs
Times GA fitness function evaluated Times GA fitness function evaluated

Fig. 12: When GA finds all six hyperparameters, the GA+DDPG+HER

training evaluation charts. This is the outcome of a single GA run.

ten runs. Table 2 compares the number of episodes an agent needs to achieve
a goal. The bolded values imply superior performance, and the majority of
Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforcement L

environments outperform the rest of the algorithms. The task is learned in

54.34% fewer episodes in the FetchPush environment than in DDPG+HER.

(a) FetchPick&Place (d) FetchSlide

(b) FetchPush (e) DoorOpening

(c) FetchReach (f) AuboReach

Fig. 13: When all six hyperparameters are determined using GA, the
DDPG+HER vs. GA+DDPG+HER efficiency evaluation plots (Total reward
vs episodes). Over ten runs, all graphs are averaged.

Another factor to consider while evaluating an algorithm’s efficiency is its

running time. Seconds are used to measure time. The algorithm is better if
24 Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforce

Method Fetch Fetch Fetch Fetch Door Aubo

Pick & Push Reach Slide Open- Reach
Place ing
DDPG+ 6,000 2,760 100 4,380 960 320
HER
GA+ 2,270 1,260 60 6,000 180 228
DDPG+
HER
PPO 2,900 2,900 1,711 4,880 1,500 2,900
A2C 119,999 119,999 119,999 119,999 119,999 32,512.3
DDPG 10,000 2,000 423 10,000 706 1,000
Table 2: Efficiency evaluation: For all activities, compare average (over ten
runs) episodes to accomplish the target.

it takes less time to learn the task. In the majority of the environments, the
GA+DDPG+HER algorithm had the lowest running time, as shown in Table
3. For example, When compared to the DDPG+HER algorithm, FetchPush
with GA+DDPG+HER takes about 57.004% less time.

Method Fetch Fetch Fetch Fetch Door Aubo

Pick & Push Reach Slide Open- Reach
Place ing
DDPG+ 3,069.981 1,314.477 47.223 2,012.645 897.816 93.258
HER
GA+ 1,224.697 565.178 28.028 3,063.599 167.883 66.818
DDPG+
HER
PPO 1,964.411 2,154.052 776.512 2,379.393 997.779 710.576
A2C 2,025.344 2,082.807 2,061.807 2,268.114 2,718.769 214.075
DDPG 5,294.984 1,000.586 236.4 5,346.516 438.7 1,721.992
Table 3: Efficiency evaluation: For all activities, compare the average (over
ten runs) running time (s) to attain the target.

The average number of steps necessary to achieve the goal is another factor
to consider when analyzing and investigating the GA+DDPG+HER algo-
rithm’s performance. The average number of steps taken by an agent in each
environment is shown in Table 4. Except for the FetchSlide environment, most
of the environments, when employed with GA+DDPG+HER, outperform all
other algorithms. When compared to the DDPG+HER algorithm, FetchPush
with GA+DDPG+HER takes about 54.35% fewer steps.
The number of epochs taken by the agent to attain the goal is the final
hyperparameter used to compare the competency of GA+DDPG+HER with
four other algorithms. The average epochs for all of the environments are
shown in Table 5. Almost all environments outperform GA+DDPG+HER
in terms of efficacy. FetchPush, for example, uses 54.35% fewer epochs with
GA+DDPG+HER than it does with DDPG+HER. Following that, we give a
comparison of the GA+DDPG+HER algorithm to the other algorithms.
Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforcement L

Method Fetch Fetch Fetch Fetch Door Aubo

Pick & Push Reach Slide Open- Reach
Place ing
DDPG+ 300,000 138,000 5,000 219,000 48,000 65,600
HER
GA+ 113,000 63,000 3,000 300,000 9,000 46,000
DDPG+
HER
PPO 595,968 595,968 324,961 1,000,000 301,056 595,968
A2C 600,000 600,000 600,000 600,000 600,000 162,566.5
DDPG 500,000 100,000 21,150 500,000 35,300 200,000
Table 4: Efficiency evaluation: For all activities, compare the average (over
ten runs) steps taken to attain the target.

Method Fetch Fetch Fetch Fetch Door Aubo

Pick & Push Reach Slide Open- Reach
Place ing
DDPG+ 60 27.6 5 43.8 47 16
HER
GA+ 22.6 12.6 3 60 8 11.4
DDPG+
HER
PPO 290 290 171.1 488 150 290
A2C 1,200 1,200 1,200 1,200 1,200 325.2
DDPG 1,000 100 42.3 1,000 70.6 100
Table 5: Efficiency evaluation: Average (over ten runs) epochs comparison to
reach the goal, for all the tasks.

Next, we present the overall analysis of the GA+DDPG+HER algorithm

compared to the other algorithms.

4.6 Analysis
We provided various findings and the mechanism for judging the efficacy of
GA+DDPG+HER versus DDPG+HER, PPO, A2C, and DDPG algorithms
in the previous sub-sections. Overall, GA+DDPG+HER works best, with
one exception of the FetchSlide environment. The average comparison tables
illustrate that different values of the assessment hyperparameters can be
assumed by each environment. This is determined by the type of task that
the agent is attempting to learn. While the majority of the tasks outper-
formed DDPG+HER with more than a 50% increase in efficiency, FetchSlide
underperformed DDPG+HER. The task’s goal is also credited with this per-
formance. The end-effector does not physically go to the desired position to
place the box, which makes this task unique. GA+DDPG+HER was tested
using a variety of hyperparameters and an average of over ten runs. This is suf-
ficient proof that GA+DDPG+HER outperformed several algorithms. Figures
5, 6 and 9b support our claim by demonstrating that when GA+DDPG+HER
26 Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforce

is utilized, the task may be learned substantially faster in most environ-

ments. Finally, in Figure 14, we compare the results of five different algorithms
(PPO, A2C, DDPG, DDPG+HER, GA+DDPG+HER). As can be shown,
GA+DDPG+HER outperforms all other algorithms in every environment
with the exception of FetchSlide, where DDPG+HER performs better than
GA+DDPG+HER. Even for the winning DDPG+HER, we think the overall
success rate on this problem (FetchSlide) is low (the highest success rate is
somewhere around 0.6 or 60%). In our opinion, GA+DDPG+HER works sim-
ilarly to its counterpart DDPG+HER in terms of problem-solving speed but
differs when an issue may have less genetic solutions. In every environment,
DDPG+HER outperforms other algorithms (such as PPO, DDPG, and A2C).
This might be because the addition of HER has improved DDPG’s perfor-
mance. Additionally, when rewards are sparse, DDPG+HER performs better
[75]. DDPG alone (without HER) cannot perform in all environments (except
in FetchReach and DoorOpening environments) as it has zero success rate in
these environments. We think this is because, for DDPG, dense rewards are
often simpler to learn from, whereas sparse rewards are more challenging. The
environment of FetchReach is undoubtedly quite simple, and all configurations
can successfully address it. DDPG is capable of performing in a DoorOpening
environment, although not as well as DDPG+HER and GA+DDPG+HER.
With the exception of FetchReach, all environments for PPO show a zero suc-
cess rate. This might be the case because PPO functions better in discrete
environments as opposed to continuous environments. This may also apply to
A2C, which has a success rate of zero in all environments, with the exception
of FetchReach and AuboReach. It can be concluded that GA+DDPG+HER
outperforms all of these techniques. This demonstrates how performance may
be improved by using GA for hyperparameter tuning.

5 Conclusion and Future Work

This study provided early data that showed how a genetic algorithm might
enhance the hyperparameters of a reinforcement learning system to enhance
performance, as demonstrated by quicker competence on six manipulation
tasks. We reviewed previous work in reinforcement learning in robotics, pre-
sented the GA+DDPG+HER algorithm for reducing the number of epochs
required to reach maximum performance, and described why a GA would be
appropriate for such optimization. Our results on the six manipulation tasks
show that the GA can identify hyperparameter values that lead to faster
learning and higher (or equal) performance at our chosen tasks, confirming
our hypothesis that GAs are a suitable fit for such hyperparameter optimiza-
tion. We compared GA+DDPG+HER to other approaches and found ours
as the most effective. In conclusion, the implementation supports the claim
that real robots learn more quickly when hyperparameters for DDPG+HER
are discovered using GA+DDPG+HER. Faster transitions between initial
and target states were possible in the real robot environment, AuboReach.
Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforcement L

(a) FetchPick&Place (d) FetchSlide

(b) FetchPush (e) DoorOpening

(c) FetchReach (f) AuboReach

Fig. 14: GA+DDPG+HER comparison with PPO [40], A2C [41],

DDPG+HER (DDPG[39] + HER[50]) and DDPG [39]. All plots are averaged
over ten runs.

Additionally, it may be claimed that GA+DDPG+HER acts similarly to its

equivalent of DDPG+HER in terms of problem-solving speed but suffers from
the same drawbacks when a problem may only have a limited number of genetic
solutions, such as in the FetchSlide environment.
28 Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforce

We also showed that heuristic search, as implemented by genetic and

other evolutionary computing techniques, is a feasible computational tool for
improving reinforcement learning and sensor odometry performance. Adap-
tive genetic algorithms can also be implemented to take on different sets of
hyperparameters during system execution. This could indicate online hyper-
parameter adjustment, which can help any system perform better, regardless
of the domain or testing environment type.

Declarations
Funding
This work is supported by the U.S. National Science Foundation (NSF) under
grants NSF-CAREER: 1846513 and NSF-PFI-TT: 1919127, and the U.S.
Department of Transportation, Office of the Assistant Secretary for Research
and Technology (USDOT/OST-R) under Grant No. 69A3551747126 through
INSPIRE University Transportation Center, and the Vingroup Joint Stock
Company and supported by Vingroup Innovation Foundation (VINIF) under
project code VINIF.2020.NCUD.DA094. The views, opinions, findings, and
conclusions reflected in this publication are solely those of the authors and do
not represent the official policy or position of the NSF, USDOT/OST-R, and
VINIF.

Conflicts of interest/Competing interests

The authors declare that there is no conflict of interest.

Code/Data availability
Open source code, and data used in the study is available at https://github.
com/aralab-unr/ga-drl-aubo-ara-lab.

Authors’ contributions
Adarsh Sehgal is the first author and primary contributor to this paper. Adarsh
did the research and wrote this manuscript. Nicholas Ward assisted Adarsh in
code execution and gathering results. Drs. Hung La and Sushil Louis advised
and overlooked this study.

Ethics approval
Not applicable

Consent to participate
All authors are full-informed about the content of this paper, and the consent
was given for participation.
Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforcement L

Consent for publication

All authors agree on the content of this paper for publication.

References
[1] Sutton, R.S., Barto, A.G., et al.: Introduction to Reinforcement Learning
vol. 135. MIT press, Cambridge (1998)

[2] Tebbe, J., Krauch, L., Gao, Y., Zell, A.: Sample-efficient reinforcement
learning in robotic table tennis. In: 2021 IEEE International Conference
on Robotics and Automation (ICRA), pp. 4171–4178 (2021). https://doi.
org/10.1109/ICRA48506.2021.9560764

[3] Xu, J., Li, B., Lu, B., Liu, Y.-H., Dou, Q., Heng, P.-A.: Surrol: An open-
source reinforcement learning centered and dvrk compatible platform for
surgical robot learning. In: 2021 IEEE/RSJ International Conference on
Intelligent Robots and Systems (IROS), pp. 1821–1828 (2021). https://
doi.org/10.1109/IROS51168.2021.9635867

[4] Chiu, Z.-Y., Richter, F., Funk, E.K., Orosco, R.K., Yip, M.C.: Biman-
ual regrasping for suture needles using reinforcement learning for rapid
motion planning. In: 2021 IEEE International Conference on Robotics
and Automation (ICRA), pp. 7737–7743 (2021). https://doi.org/10.1109/
ICRA48506.2021.9561673

[5] Marchesini, E., Corsi, D., Farinelli, A.: Benchmarking safe deep reinforce-
ment learning in aquatic navigation. In: 2021 IEEE/RSJ International
Conference on Intelligent Robots and Systems (IROS), pp. 5590–5595
(2021). https://doi.org/10.1109/IROS51168.2021.9635925

[6] Probst, P., Boulesteix, A.-L., Bischl, B.: Tunability: Importance of hyper-
parameters of machine learning algorithms. The Journal of Machine
Learning Research 20(1), 1934–1965 (2019)

[7] Zhang, B., Rajan, R., Pineda, L., Lambert, N., Biedenkapp, A., Chua, K.,
Hutter, F., Calandra, R.: On the importance of hyperparameter optimiza-
tion for model-based reinforcement learning. In: International Conference
on Artificial Intelligence and Statistics, pp. 4015–4023 (2021). PMLR

[8] Sebag, M., Schoenauer, M.: A society of hill-climbers. In: Proceedings

of 1997 IEEE International Conference on Evolutionary Computation
(ICEC’97), pp. 319–324 (1997). IEEE

[9] Davis, L.: Handbook of genetic algorithms (1991)

[10] Nguyen, A.-T., Reiter, S., Rigo, P.: A review on simulation-based opti-
mization methods applied to building performance analysis. Applied
30 Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforce

energy 113, 1043–1058 (2014)

[11] Deb, K.: Multi-objective genetic algorithms: Problem difficulties and

construction of test problems. Evolutionary computation 7(3), 205–230
(1999)

[12] Wetter, M., Wright, J.: A comparison of deterministic and probabilis-

tic optimization algorithms for nonsmooth simulation-based optimization.
Building and environment 39(8), 989–999 (2004)

[13] Sehgal, A., La, H., Louis, S., Nguyen, H.: Deep reinforcement learn-
ing using genetic algorithm for parameter optimization. In: 2019 Third
IEEE International Conference on Robotic Computing (IRC), pp. 596–601
(2019). IEEE

[14] Sehgal, A.: Genetic algorithm as function optimizer in reinforcement

learning and sensor odometry. Master’s thesis, University of Nevada, Reno
(2019)

[15] Sehgal, A., Ward, N., La, H.M., Papachristos, C., Louis, S.: Ga-drl:
Genetic algorithm-based function optimizer in deep reinforcement learn-
ing for robotic manipulation tasks. arXiv preprint arXiv:2203.00141
(2022)

[16] Sehgal, A., Singandhupe, A., La, H.M., Tavakkoli, A., Louis, S.J.:
Lidar-monocular visual odometry with genetic algorithm for parameter
optimization. In: Bebis, G., et al.(eds.) Advances in Visual Computing,
pp. 358–370. Springer, Cham (2019)

[17] Dhariwal, P., Hesse, C., Klimov, O., Nichol, A., Plappert, M., Radford, A.,
Schulman, J., Sidor, S., Wu, Y., Zhokhov, P.: OpenAI Baselines. GitHub
(2017)

[18] Holland, J.H.: Genetic algorithms. Scientific american 267(1), 66–73

(1992)

[19] Goldberg, D.E., Holland, J.H.: Genetic algorithms and machine learning.
Machine learning 3(2), 95–99 (1988)

[20] De Jong, K.: Learning with genetic algorithms: An overview. Machine

learning 3(2-3), 121–138 (1988)

[21] Goldberg, D.E., Deb, K.: A comparative analysis of selection schemes

used in genetic algorithms. In: Foundations of Genetic Algorithms vol. 1,
pp. 69–93. Elsevier, Netherlands (1991)

[22] Syswerda, G.: Uniform crossover in genetic algorithms. In: Proceedings

Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforcement L

of the Third International Conference on Genetic Algorithms, pp. 2–9

(1989). Morgan Kaufmann Publishers

[23] Sehgal, A., Singandhupe, A., La, H.M., Tavakkoli, A., Louis, S.J.:
Lidar-monocular visual odometry with genetic algorithm for parameter
optimization. In: International Symposium on Visual Computing, pp.
358–370 (2019). Springer

[24] Watkins, C.J., Dayan, P.: Q-learning. Machine learning 8(3-4), 279–292
(1992)

[25] La, H.M., Lim, R., Sheng, W.: Multirobot cooperative learning for preda-
tor avoidance. IEEE Transactions on Control Systems Technology 23(1),
52–63 (2015). https://doi.org/10.1109/TCST.2014.2312392

[26] Gaskett, C., Wettergreen, D., Zelinsky, A.: Q-learning in continuous

state and action spaces. In: Australasian Joint Conference on Artificial
Intelligence, pp. 417–428 (1999). Springer

[27] Doya, K.: Reinforcement learning in continuous time and space. Neural
computation 12(1), 219–245 (2000)

[28] Hasselt, H.V., Wiering, M.A.: Reinforcement learning in continuous action

spaces (2007)

[29] Baird, L.C.: Reinforcement learning in continuous time: Advantage updat-

ing. In: Neural Networks, 1994. IEEE World Congress on Computational
Intelligence., 1994 IEEE International Conference On, vol. 4, pp. 2448–
2453 (1994). IEEE

[30] Wei, Q., Lewis, F.L., Sun, Q., Yan, P., Song, R.: Discrete-time deter-
ministic q-learning: A novel convergence analysis. IEEE transactions on
cybernetics 47(5), 1224–1237 (2017)

[31] Kohl, N., Stone, P.: Policy gradient reinforcement learning for fast
quadrupedal locomotion. In: Robotics and Automation, 2004. Proceed-
ings. ICRA’04. 2004 IEEE International Conference On, vol. 3, pp.
2619–2624 (2004). IEEE

[32] Endo, G., Morimoto, J., Matsubara, T., Nakanishi, J., Cheng, G.:
Learning cpg-based biped locomotion with a policy gradient method:
Application to a humanoid robot. The International Journal of Robotics
Research 27(2), 213–228 (2008)

[33] Peters, J., Mülling, K., Altun, Y.: Relative entropy policy search. In:
AAAI, pp. 1607–1612 (2010). Atlanta
32 Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforce

[34] Kalakrishnan, M., Righetti, L., Pastor, P., Schaal, S.: Learning force
control policies for compliant manipulation. In: Intelligent Robots and
Systems (IROS), 2011 IEEE/RSJ International Conference On, pp.
4639–4644 (2011). IEEE

[35] Vargas, J.C., Bhoite, M., Barati Farimani, A.: Creativity in robot manip-
ulation with deep reinforcement learning. arXiv e-prints, 1910 (2019)

[36] Kober, J., Bagnell, J.A., Peters, J.: Reinforcement learning in robotics: A
survey. The International Journal of Robotics Research 32(11), 1238–1274
(2013)

[37] Andrew, A.M.: Reinforcement learning: An introduction by richard s. sut-

ton and andrew g. barto, adaptive computation and machine learning
series, mit press (bradford book), cambridge, mass., 1998, xviii+ 322 pp,
isbn 0-262-19398-1,(hardback,£ 31.95). Robotica 17(2), 229–235 (1999)

[38] Munos, R., Stepleton, T., Harutyunyan, A., Bellemare, M.: Safe and
efficient off-policy reinforcement learning. In: Advances in Neural Infor-
mation Processing Systems, pp. 1054–1062 (2016)

[39] Lillicrap, T.P., Hunt, J.J., Pritzel, A., Heess, N., Erez, T., Tassa, Y.,
Silver, D., Wierstra, D.: Continuous control with deep reinforcement
learning. arXiv preprint arXiv:1509.02971 (2015)

[40] Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal
policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)

[41] Mnih, V., Badia, A.P., Mirza, M., Graves, A., Lillicrap, T., Harley, T.,
Silver, D., Kavukcuoglu, K.: Asynchronous methods for deep reinforce-
ment learning. In: International Conference on Machine Learning, pp.
1928–1937 (2016). PMLR

[42] Gu, S., Lillicrap, T., Sutskever, I., Levine, S.: Continuous deep q-learning
with model-based acceleration. In: International Conference on Machine
Learning, pp. 2829–2838 (2016)

[43] Deisenroth, M.P., Rasmussen, C.E., Fox, D.: Learning to control a low-
cost manipulator using data-efficient reinforcement learning (2011)

[44] Jin, L., Li, S., La, H.M., Luo, X.: Manipulability optimization of redun-
dant manipulators using dynamic neural networks. IEEE Transactions on
Industrial Electronics 64(6), 4710–4720 (2017). https://doi.org/10.1109/
TIE.2017.2674624

[45] Lin, C.-K.: H∞ reinforcement learning control of robot manipulators

using fuzzy wavelet networks. Fuzzy Sets and Systems 160(12), 1765–1786
Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforcement L

(2009)

[46] Miljković, Z., Mitić, M., Lazarević, M., Babić, B.: Neural network rein-
forcement learning for visual control of robot manipulators. Expert
Systems with Applications 40(5), 1721–1736 (2013)

[47] Duguleana, M., Barbuceanu, F.G., Teirelbar, A., Mogan, G.: Obstacle
avoidance of redundant manipulators using neural networks based rein-
forcement learning. Robotics and Computer-Integrated Manufacturing
28(2), 132–146 (2012)

[48] Nguyen, H., La, H.M.: Review of deep reinforcement learning for robot
manipulation. In: 2019 Third IEEE International Conference on Robotic
Computing (IRC), pp. 590–595 (2019). IEEE

[49] Chane-Sane, E., Schmid, C., Laptev, I.: Goal-conditioned reinforcement

learning with imagined subgoals. In: International Conference on Machine
Learning, pp. 1430–1440 (2021). PMLR

[50] Andrychowicz, M., Wolski, F., Ray, A., Schneider, J., Fong, R., Welinder,
P., McGrew, B., Tobin, J., Abbeel, O.P., Zaremba, W.: Hindsight experi-
ence replay. In: Advances in Neural Information Processing Systems, pp.
5048–5058 (2017)

[51] Nguyen, H., La, H.M., Deans, M.: Deep learning with experience ranking
convolutional neural network for robot manipulator. arXiv:1809.05819,
cs.RO (2018) https://arxiv.org/abs/1809.05819

[52] Pham, H.X., La, H.M., Feil-Seifer, D., Nguyen, L.V.: Autonomous uav
navigation using reinforcement learning. arXiv:1801.05086, cs.RO (2018)
https://arxiv.org/abs/1801.05086

[53] Pham, H.X., La, H.M., Feil-Seifer, D., Nguyen, L.V.: Reinforcement learn-
ing for autonomous uav navigation using function approximation. In: 2018
IEEE International Symposium on Safety, Security, and Rescue Robotics
(SSRR), pp. 1–6 (2018). https://doi.org/10.1109/SSRR.2018.8468611

[54] La, H.M., Lim, R.S., Sheng, W., Chen, J.: Cooperative flocking and
learning in multi-robot systems for predator avoidance. In: 2013 IEEE
International Conference on Cyber Technology in Automation, Control
and Intelligent Systems, pp. 337–342 (2013). https://doi.org/10.1109/
CYBER.2013.6705469

[55] La, H.M., Sheng, W., Chen, J.: Cooperative and active sensing in mobile
sensor networks for scalar field mapping. IEEE Transactions on Systems,
Man, and Cybernetics: Systems 45(1), 1–12 (2015). https://doi.org/10.
1109/TSMC.2014.2318282
34 Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforce

[56] Pham, H.X., La, H.M., Feil-Seifer, D., Nefian, A.: Cooperative
and distributed reinforcement learning of drones for field coverage.
arXiv:1803.07250, cs.RO (2018) https://arxiv.org/abs/1803.07250

[57] Dang, A.D., La, H.M., Horn, J.: Distributed formation control for
autonomous robots following desired shapes in noisy environment. In:
2016 IEEE International Conference on Multisensor Fusion and Integra-
tion for Intelligent Systems (MFI), pp. 285–290 (2016). https://doi.org/
10.1109/MFI.2016.7849502

[58] Rahimi, M., Gibb, S., Shen, Y., La, H.M.: A comparison of vari-
ous approaches to reinforcement learning algorithms for multi-robot
box pushing. In: International Conference on Engineering Research and
Applications, pp. 16–30 (2018). Springer

[59] Deb, K., Pratap, A., Agarwal, S., Meyarivan, T.: A fast and elitist mul-
tiobjective genetic algorithm: Nsga-ii. IEEE transactions on evolutionary
computation 6(2), 182–197 (2002)

[60] Poon, P.W., Carter, J.N.: Genetic algorithm crossover operators for
ordering applications. Computers & Operations Research 22(1), 135–147
(1995)

[61] Liu, F., Zeng, G.: Study of genetic algorithm with reinforcement learn-
ing to solve the tsp. Expert Systems with Applications 36(3), 6995–7001
(2009)

[62] Moriarty, D.E., Schultz, A.C., Grefenstette, J.J.: Evolutionary algorithms

for reinforcement learning. Journal of Artificial Intelligence Research 11,
241–276 (1999)

[63] Mikami, S., Kakazu, Y.: Genetic reinforcement learning for cooperative
traffic signal control. In: Evolutionary Computation, 1994. IEEE World
Congress on Computational Intelligence., Proceedings of the First IEEE
Conference On, pp. 223–228 (1994). IEEE

[64] Sutton, R.S., McAllester, D., Singh, S., Mansour, Y.: Policy gradi-
ent methods for reinforcement learning with function approximation.
Advances in neural information processing systems 12 (1999)

[65] Alagha, A., Singh, S., Mizouni, R., Bentahar, J., Otrok, H.: Target
localization using multi-agent deep reinforcement learning with proximal
policy optimization. Future Generation Computer Systems 136, 342–357
(2022)

[66] Wu, C.-A.: Investigation of Different Observation and Action Spaces for
Reinforcement Learning on Reaching Tasks (2019)
Automatic Hyperparameter Optimization Using Genetic Algorithm in Deep Reinforcement L

[67] Melo, L.C., Máximo, M.R.O.A.: Learning humanoid robot running skills
through proximal policy optimization. In: 2019 Latin American Robotics
Symposium (LARS), 2019 Brazilian Symposium on Robotics (SBR) and
2019 Workshop on Robotics in Education (WRE), pp. 37–42 (2019).
IEEE

[68] Hsu, C.C.-Y., Hardt, C.M.-D.M.: Revisiting design choices in proximal

policy optimization. arXiv preprint arXiv:2009.10897 (2020)

[69] Scholte, N.: Goal, mistake and success learning through resimulation.
Master’s thesis (2022)

[70] Haydari, A., Zhang, M., Chuah, C.-N., Ghosal, D.: Impact of deep rl-
based traffic signal control on air quality. In: 2021 IEEE 93rd Vehicular
Technology Conference (VTC2021-Spring), pp. 1–6 (2021). IEEE

[71] Polyak, B.T., Juditsky, A.B.: Acceleration of stochastic approximation by

averaging. SIAM Journal on Control and Optimization 30(4), 838–855
(1992)

[72] Brockman, G., Cheung, V., Pettersson, L., Schneider, J., Schulman, J.,
Tang, J., Zaremba, W.: Openai gym. arXiv preprint arXiv:1606.01540
(2016)

[73] Hernandez-Mendez, S., Maldonado-Mendez, C., Marin-Hernandez, A.,

Rios-Figueroa, H.V., Vazquez-Leal, H., Palacios-Hernandez, E.R.: Design
and implementation of a robotic arm using ros and moveit! In: 2017 IEEE
International Autumn Meeting on Power, Electronics and Computing
(ROPEC), pp. 1–6 (2017). IEEE

[74] Field, G., Stepanenko, Y.: Iterative dynamic programming: an approach

to minimum energy trajectory planning for robotic manipulators. In: Pro-
ceedings of IEEE International Conference on Robotics and Automation,
vol. 3, pp. 2755–2760 (1996). IEEE

[75] Plappert, M., Andrychowicz, M., Ray, A., McGrew, B., Baker, B., Powell,
G., Schneider, J., Tobin, J., Chociej, M., Welinder, P., et al.: Multi-goal
reinforcement learning: Challenging robotics environments and request for
research. arXiv preprint arXiv:1802.09464 (2018)