The Decision Transformer model, introduced by Chen L. et al. in “Decision Transformer: Reinforcement Learning via Sequence Modeling,” transforms the reinforcement learning (RL) landscape by treating RL as a conditional sequence modeling problem.

1. 1/22
Decision Transformers Model
leewayhertz.com/decision-transformer
The realm of Reinforcement Learning (RL) has experienced a surge of innovations in recent
years, propelling artificial intelligence towards its maximum potential. One such advancement
is the Decision Transformer, a model that melds the power of Transformer architectures with
the adaptability of reinforcement learning. The emergence of Decision Transformers marks a
crucial milestone in the evolution of machine learning, showcasing their remarkable potential
in transforming how RL-based agents interact with their environments.
Decision Transformers harness the strengths of Transformer models, known for their
prowess in handling sequential data such as natural language processing, and couple them
with the dynamic learning capabilities of reinforcement learning. This convergence allows for
a more efficient and effective way to train intelligent agents, overcoming traditional RL
methods’ limitations.
By leveraging the power of transformers, Decision Transformers enable offline reinforcement
learning, reducing the need for resource-intensive online training, enabling agents to learn
from existing datasets. This accelerates learning and mitigates risks associated with training
agents in real-world environments or flawed simulators.
Furthermore, Decision Transformers address long-term dependencies, a persistent
challenge in RL, by handling complex sequential data and generating future action
sequences to optimize reward outcomes. This innovative approach has far-reaching

2. 2/22
implications for a wide range of applications, from robotics and autonomous vehicles to
strategic gameplay and personalized user experiences.
By embracing the synergy between Transformer architectures and reinforcement learning,
Decision Transformers pave the way for a new era of AI-powered solutions and
advancements.
This article delves into the evolution of the Transformer model and highlights the challenges
in reinforcement learning. It introduces offline reinforcement learning and explains sequence
modeling via RL. The Decision Transformer model is presented with a detailed overview of
its architecture and its key components. The article also explains how the Decision
Transformer works and outlines its various applications and use cases. Lastly, the article
offers guidance on how to use a Decision Transformer in a Transformer model, providing
valuable insights for anyone interested in this innovative technology.
Evolution of Transformer-based models
Challenges of reinforcement learning
What is offline reinforcement learning?
What is reinforcement learning via sequence modeling?
What is the Decision Transformer model? Its architecture and key components
How does the Decision Transformer work?
Applications and use cases of the Decision Transformer
How to use the Decision Transformer in a Transformer?
Evolution of Transformer-based models
This section will briefly discuss the original Transformer architecture and four significant
Transformer architectures that emerged shortly after its introduction.
Before diving in, let’s understand the kind of tasks various models perform.
Encoder-only models: Suitable for tasks requiring input comprehension, such as
sentiment analysis.
Decoder-only models: Designed for tasks that involve text generation, like story-
writing.
Encoder-decoder models or sequence-to-sequence models: Used for tasks that
require generating text from input, like summarization.
Transformer by Google
In 2017, Google researchers introduced a network architecture called Transformer in their
paper “Attention is all you need.” Designed primarily for natural language processing tasks,
Transformers changed the field by addressing the limitations of Recurrent Neural Networks
(RNNs).

3. 3/22
Unlike RNNs, which handle one word at a time, Transformers can process complete
sentences at once. This makes Transformers more efficient and better at understanding
context. The attention mechanism helps Transformers access past information more
accurately than RNNs, which have difficulty keeping context beyond the previous state.
Transformers also use positional embeddings to track word positions within sentences. This
advantage has led Transformers to outperform RNNs in almost all language tasks.
Undoubtedly, learning about Transformers’ innovative design and incredible growth in recent
years is essential.
This groundbreaking architecture led to the emergence of four significant Transformer-based
models: OpenAI’s GPT, Google’s BERT, OpenAI’s GPT-2, and Facebook’s RoBERTa.
GPT by OpenAI
Output: Probabilities Over Tokens
Softmax
Add & Layer Norm
Pointwise Feed Forward
Add & Layer Norm
Masked Multi-Headed Self-Attention
Input: X
hL
We
T
hL
xWe
+Wp
Embedding Matrix We
Transposed Embedding We
T
Transformer Block
Repeat xL = 12
Transformer_block
1,…,L
hl =
(hl
-1)
l =
LeewayHertz
The evolution of GPT in recent years is nothing short of remarkable, making it an excellent
starting point for discussing Transformers. GPT, short for Generative Pre-training, pioneered
the concept of unsupervised learning for pre-training, followed by supervised learning for
fine-tuning—a method now widely adopted by many Transformers.

4. 4/22
GPT was initially trained on a unique dataset comprising 7,000 unpublished books,
showcasing its prowess in understanding complex language patterns. Its architecture,
composed of 12 stacked decoders, distinguishes it as a decoder-only model. Despite having
117 million parameters, GPT-1’s capacity pales compared to the more advanced
Transformers developed today. GPT’s unidirectional nature, which involves masking tokens
to the right of the current token, allows it to generate one token at a time and use that output
for the subsequent timestep.
GPT marked the dawn of a new era for Transformers, paving the way for even more
astonishing innovations that would soon follow.
The evolution of GPT models is a testament to the incredible advancements in natural
language processing. Spanning several iterations, GPT models have consistently pushed the
boundaries of what AI language models can achieve.
GPT (GPT-1): The first iteration, GPT, or Generative Pre-trained Transformer,
introduced the concept of unsupervised learning for pre-training and supervised fine-
tuning. It utilized 12 decoders in its architecture and was trained on a dataset of 7,000
unpublished books. With 117 million parameters, GPT-1 was a groundbreaking model
for its time, though its successors have since eclipsed it.
GPT-2: Building upon the foundation of GPT-1, GPT-2 featured an impressive 1.5
billion parameters, significantly increasing the model’s capabilities. It utilized 48
decoders in its architecture and was trained on a much larger dataset of 8 million web
pages. Despite concerns about potential misuse, GPT-2 demonstrated an extraordinary
ability to generate coherent and contextually relevant text.
GPT-3: The latest powerful iteration, GPT-3, boasts 175 billion parameters. This model
has shown remarkable performance across a wide range of tasks, including translation,
summarization, and even programming assistance. Its advanced architecture and
unparalleled capacity for contextual understanding have positioned GPT-3 as a leader
in the field of natural language processing.

5. 5/22
GPT-4: It is the most recent breakthrough in OpenAI’s continuous endeavor to scale up
deep learning. As a multimodal model, GPT-4 is designed to accept both image and
text inputs while generating text outputs. This innovative approach allows the model to
demonstrate human-level performance across numerous professional and academic
benchmarks. GPT-4 builds on the success of its predecessors, incorporating the
advancements made in earlier iterations while pushing the boundaries of AI
capabilities. Its ability to process and understand both visual and textual data enables
GPT-4 to tackle a wider range of tasks and adapt to diverse real-world applications
more effectively. While GPT-4 may not yet surpass human performance in every
scenario, it represents a significant step forward in the field of artificial intelligence.
GPT-4 demonstrates its potential to revolutionize industries and reshape our
understanding of AI’s role in modern society by excelling in various benchmarks. With
the release of GPT-4, OpenAI reaffirms its commitment to advancing deep learning and
creating models that improve upon existing technology and pave the way for future
breakthroughs. As the AI community continues to explore the possibilities offered by
GPT-4, we can expect further innovations and exciting developments in the realm of
artificial intelligence.
Google’s BERT

6. 6/22
W3
Classification Layer: Fully-Connected Layer + GELU + Norm
Transformer Encoder
Embedding
Embedding
To Vocal +
Softmax
W’ 1
W1
W1
O1
W2
W2
W’ 2
O2
W3
W’ 3
O3
W4
[MASK]
W’ 4
O4
W5
W5
W’ 5
O5
LeewayHertz
BERT, the Bidirectional Encoder Representation from Transformers, is a game-changing
model that redefines how we approach natural language processing. BERT’s bidirectional
nature sets it apart from its counterparts, as its attention mechanism can effectively capture
the context from both the left and right directions of the current token. With a powerful
architecture comprising 12 stacked encoders, BERT serves as an encoder-only model that
can process entire sentences as input, allowing it to reference any word in the sentence for
unparalleled performance.
Boasting 110 million parameters, BERT shares the flexibility of GPT, as it can be trained on
specific tasks and fine-tuned for a variety of applications. Its unique pre-training method,
which essentially involves a fill-in-the-blank task, demonstrates the exceptional capabilities of
this highly efficient model.
Facebook’s RoBERTa
Facebook’s RoBERTa, Robustly Optimized BERT Pre-training Approach, takes the power of
BERT several notches higher. While RoBERTa shares the same architecture as BERT, the
significant optimization it has undergone sets it apart. These enhancements have led to

7. 7/22
state-of-the-art results on various NLP benchmarks, including MNLI, QNLI, RTE, STS-B,
GLUE, and RACE tasks.
By eliminating BERT’s next-sentence pretraining objective, training with larger mini-batches
and altering the masking pattern, RoBERTa has surpassed its predecessor in performance.
Furthermore, RoBERTa has been trained on a more extensive dataset, incorporating existing
unannotated NLP datasets and CC-News, a novel collection of public news articles.
The rapid evolution of Transformer-based models has significantly impacted the field of
natural language processing. Researchers have created new models that perform specific
NLP tasks more effectively than the original Transformer by separating and stacking encoder
and decoder architectures. The widespread adoption of unsupervised learning for pre-
training and supervised learning for fine-tuning has furthered the development of these
models.
Challenges of reinforcement learning
Reinforcement learning is a branch of machine learning where agents learn to make
decisions in an environment by interacting with it and receiving feedback in the form of
rewards. The goal is to maximize the cumulative reward over time by discovering the best
sequence of actions. This learning process enables agents to adapt their behavior and
improve their decision-making capabilities in various situations. Although widely employed in
applications such as autonomous driving, robot navigation and game playing, RL has several
significant limitations, such as:
Sample efficiency
Sample efficiency is a challenge in reinforcement learning (RL) as it necessitates numerous
interactions with the environment to learn an effective policy. Real-world scenarios like
robotics or autonomous driving make these interactions time-consuming, costly, or risky.
Efficient exploration strategies are vital for RL agents to gather valuable experiences while
navigating their environment. Complex environments with vast state and action spaces
require a larger number of samples to find optimal policies. Additionally, limited sample
efficiency can hinder generalization, leading to poor performance in novel situations.
Exploration vs. exploitation
This is a limitation of reinforcement learning because it represents a fundamental trade-off
that RL agents must navigate during the learning process. Exploration involves trying new
actions in different states to gather information about the environment, while exploitation
leverages the agent’s current knowledge to make decisions that maximize the reward
function. Striking the right balance is crucial, as focusing too much on exploration may lead
to suboptimal decisions and lower cumulative rewards, while excessive exploitation may

8. 8/22
cause the agent to miss out on discovering potentially better strategies. Achieving this
balance is particularly challenging in complex and dynamic environments where the optimal
policy may change over time. Thus, developing RL algorithms that effectively manage the
exploration-exploitation trade-off is essential for ensuring their success and applicability in
real-world tasks.
Sparse rewards
Sparse rewards represent a limitation of reinforcement learning because they make it difficult
for agents to identify and learn from informative feedback. When rare non-zero rewards are
only provided under particular circumstances, the agent may struggle to discern the causal
relationship between its actions and the corresponding rewards. As a result, the learning
process can become slow and inefficient. In environments with sparse rewards, the agent
might need to take a long sequence of actions before receiving any meaningful feedback,
making it harder to pinpoint which actions contributed to achieving the reward. Consequently,
this may lead to extended training times or suboptimal performance, as the agent could fail
to discover the most effective strategies. Researchers have developed various techniques to
address the challenge of sparse rewards, such as reward shaping, hierarchical
reinforcement learning, and intrinsic motivation, to provide more informative feedback or
structure the learning process to guide the agent more effectively toward the optimal policy.
Long-term dependencies
Long-term dependencies represent a significant challenge in reinforcement learning because
they require agents to consider the consequences of their actions over extended time
horizons. In tasks like chess or Go, an agent must be able to plan and strategize multiple
steps ahead, considering the potential outcomes of each move and their subsequent impact
on the game’s state. When dealing with long-term dependencies, it becomes difficult for the
agent to attribute the final reward or outcome to specific actions taken earlier in the
sequence. This issue, known as the credit assignment problem, can slow down the learning
process and hinder the agent’s ability to discover optimal policies. Several approaches have
been developed to address the challenge of long-term dependencies in reinforcement
learning. Techniques such as temporal-difference learning, eligibility traces, and Monte Carlo
tree search methods have been used to help the agent better understand the long-term
consequences of its actions and improve its decision-making capabilities in tasks with
extended time horizons.
Stochastic environments
Stochastic environments pose a significant challenge for reinforcement learning because
they introduce randomness and uncertainty into the dynamics of the system. In such
environments, the outcomes of actions are not deterministic, meaning that the same action

9. 9/22
may lead to different results when applied in the same state. This unpredictability can make it
difficult for RL agents to learn effective policies and strategies.
Learning in stochastic environments requires the agent to cope with uncertainty and estimate
the probability distributions of different outcomes. This adds complexity to the learning
process and can slow down convergence to the optimal policy. To deal with these
challenges, RL algorithms need to be robust and capable of adapting to the inherent
variability in the environment.
Various techniques have been developed to address stochastic environments’ challenges in
reinforcement learning. These include methods such as Bayesian reinforcement learning,
which explicitly models environmental uncertainty and algorithms incorporating exploration
strategies to gather more information about the environment’s dynamics. By incorporating
these approaches, RL agents can better learn to navigate and make decisions in stochastic
environments despite the inherent randomness and uncertainty.
Transfer learning
Transfer learning is a limitation of reinforcement learning because it involves leveraging
knowledge acquired from one task or environment and applying it to a different but related
task or environment. The goal is to improve learning efficiency and performance in the target
task by reusing the knowledge from the source task. However, RL algorithms often struggle
to achieve this effectively, which can limit their practical applications.
There are several reasons for compromised transfer learning capabilities in RL, such as:
Domain differences: The source and target tasks may have distinct state and action
spaces, making it challenging to directly apply the knowledge from one to the other.
Task variations: The reward functions and dynamics of the source and target tasks
may differ, complicating the process of transferring knowledge.
Negative transfer: Knowledge transfer between tasks can sometimes decrease
performance if the source and target tasks are not sufficiently similar.
Safety and reliability
Safety and reliability are critical concerns in reinforcement learning (RL), particularly when
dealing with high-stakes or real-world applications. Ensuring an RL agent makes safe and
reliable decisions is crucial to prevent unintended consequences, accidents, or failures.
However, achieving safety and reliability presents several challenges:
Exploration risks: RL agents explore different actions to discover the optimal policy
during learning. This exploration can lead to actions that might have negative
consequences, such as damaging equipment or causing harm in a physical
environment.

10. 10/22
Lack of performance guarantees: Traditional RL algorithms often lack performance
guarantees during the learning process, which makes it difficult to ensure that the
agent will always behave safely and reliably.
Partial observability: In many real-world environments, the agent may only have
access to partial information about the system’s state, making it challenging to
determine safe actions with certainty.
Uncertainty and noise: Real-world environments often contain uncertainty and noise,
such as sensor errors or unpredictable environmental changes. This uncertainty can
make it difficult to accurately estimate the true state of the environment and its
dynamics, leading to potential safety and reliability issues.
Distributional shift: When an RL agent is deployed in a new environment, it may
encounter situations that were not present in its training data. This distributional shift
can result in the agent making suboptimal or unsafe decisions based on its previous
experience.
Overcoming these challenges is essential for advancements in the field of reinforcement
learning and realizing its full potential across applications.
What is offline reinforcement learning?
Offline reinforcement learning, also known as batch reinforcement learning or data-driven
reinforcement learning, is a method of training decision-making agents without direct
interaction with the environment. Instead, agents learn from pre-collected data, which can
originate from other agents or human demonstrations. This approach offers an alternative to
online reinforcement learning, where agents gather data and learn from it through direct
interaction with the environment or a simulator.

11. 11/22
Environment
State,Reward
Action
ReinforcementLearningwithOnlineInteractions
State,Actions,Rewards
Environment
Logged
Interaction
OfflineReinforcementLearning
OnlineAgent
OfflineAgent
LeewayHertz
Offline reinforcement learning involves the following process:
Create a dataset using one or more policies and/or human interactions.
Run offline RL on this dataset to learn a policy.
This method is particularly beneficial when building a complex, expensive, or insecure
simulator. However, offline reinforcement learning has a major drawback known as the
counterfactual queries problem. This issue arises when an agent decides to perform an
action for which no data is available in the pre-collected dataset. For example, suppose the
agent wants to turn right at an intersection but the dataset does not contain information about
this trajectory. In that case, the agent may struggle to make an informed decision.
Despite this limitation, offline reinforcement learning offers a valuable alternative for training
deep reinforcement learning agents, especially in situations where real-world interactions or
simulator construction is challenging or impractical.

12. 12/22
What is reinforcement learning via sequence modeling?
Reinforcement learning via sequence modeling is a method that combines reinforcement
learning with sequence modeling techniques to address problems where the agent’s
decision-making process involves a sequence of actions or states with temporal
dependencies. This approach is necessary when the agent needs to learn optimal strategies
in complex environments where the relationships between actions and states evolve over
time.
Sequence modeling is a technique in machine learning that focuses on analyzing and
predicting sequences of data points. It deals with problems where the order and relationships
between data points are significant, such as time series forecasting, natural language
processing, and protein sequence analysis. Common algorithms used in sequence modeling
include recurrent neural networks (RNNs), long short-term memory (LSTM) networks, and
Transformers, which can capture complex patterns and dependencies within the data
sequences. By understanding these patterns, sequence modeling techniques can generate
predictions or generate entirely new sequences based on the learned context. These
techniques allow the agent to learn from the history of states, actions, and rewards, which in
turn helps it generate an optimal sequence of actions that maximize the cumulative reward
over time.
The necessity of reinforcement learning via sequence modeling arises in various
applications, including:
Natural language understanding: In tasks like dialogue systems or machine
translation, agents must generate appropriate responses based on the context
provided by the input text, which often requires understanding the relationships
between words and phrases over time.
Time-series prediction: When predicting future values in time-series data, such as
stock prices or weather patterns, it is crucial to learn the data’s temporal dependencies
and underlying patterns, which can be addressed using sequence modeling techniques
combined with reinforcement learning.
Robot control: In robotics, agents often need to learn the optimal sequence of actions
to complete specific tasks, such as object manipulation or navigation, based on the
history of states and rewards. Reinforcement learning via sequence modeling provides
a suitable framework to tackle such problems.
What is the Decision Transformer model? Its architecture and key
components
The Decision Transformer model, introduced by Chen L. et al. in “Decision Transformer:
Reinforcement Learning via Sequence Modeling,” transforms the reinforcement learning (RL)
landscape by treating RL as a conditional sequence modeling problem. Instead of relying on

13. 13/22
traditional RL methods, like fitting a value function to guide action selection and maximize
returns, the Decision Transformer utilizes a sequence modeling algorithm (i.e., the
Transformer) to generate future actions that achieve a specified desired return.
This innovative approach hinges on an autoregressive model conditioned on the desired
return, past states, and actions. By using generative trajectory modeling, that predicts the
combined patterns of situations, actions, and rewards, the method simplifies and makes it
easier for people to understand the reinforcement learning process, the Decision
Transformer sidesteps the conventional RL process of return maximization and directly
generates a series of future actions that fulfill the desired return.
a
t-1
a t
CausalTransformer
RTG Action State
t-1
R t-1
S t-1
a t
R t
S t
a
LinearDecoder
emb.+pos.enc.
LeewayHertz
The process is as follows:
The Decision Transformer takes information from the most recent events (the last K
timesteps) and uses three pieces of data for each event: the expected future rewards
(return-to-go), the current situation (state), and the action taken. Using this information,
the model can learn to generate actions that help achieve a desired outcome.
Tokens are embedded using a linear layer for vector states or a CNN encoder for
frame-based states. This means that the information (tokens) about the current
situation is transformed into a more useful format for the model. If the information is a
simple list of numbers (vector state), a basic method (linear layer) is used. If the
information is a series of images (frame-based states), a more complex method (CNN
encoder) designed for image processing is utilized.
These inputs are then processed by a GPT-2 model, which predicts future actions
through autoregressive modeling, employing a causal self-attention mask.

14. 14/22
The Decision Transformer model paves the way for a new era of advanced and efficient AI
systems by reimagining the RL paradigm and leveraging the power of Transformers. This
groundbreaking approach can potentially reshape reinforcement learning and unlock exciting
new applications across diverse domains.
The key components of a Decision Transformer include:
Transformer architecture: The Decision Transformer utilizes the Transformer
architecture, which is known for effectively processing sequential data. Transformers
consist of multi-headed self-attention mechanisms and feed-forward neural networks,
which enable the model to capture complex dependencies within the input data.
State-Action-Reward sequences: The Decision Transformer learns from a dataset of
state-action-reward (SAR) sequences, where each sequence represents the agent’s
interaction with the environment. These sequences are used to train the Transformer in
an offline reinforcement learning setting.
Reward modeling: The Decision Transformer uses the reward information embedded
within the SAR sequences to predict the optimal actions that lead to the highest
cumulative rewards. This reward-driven approach ensures that the generated action
sequences align with the agent’s goal of maximizing its rewards in the environment.
Context conditioning: To predict the optimal action, the Decision Transformer
conditions its predictions on the current state and the history of past states, actions,
and rewards. This context conditioning allows the model to learn the relationships
between past experiences and future actions, thus guiding the agent towards better
decisions.
Offline Reinforcement Learning: The Decision Transformer operates in an offline
reinforcement learning setting, where the agent learns from a pre-collected dataset of
SAR sequences without actively exploring the environment. This approach reduces
computational requirements and enables learning from diverse sources, such as
human demonstrations or other agents’ experiences.
Sequence generation: Leveraging the Transformer’s ability to generate future
sequences, the Decision Transformer predicts the sequence of actions that will
maximize the agent’s cumulative reward. This sequence generation capability is key to
the model’s success in reinforcement learning tasks.
Combining these components, the Decision Transformer creates a powerful model bridging
the gap between Transformers and reinforcement learning, enabling more efficient and
versatile learning in various domains.
The architecture of the Decision Transformer model
The architecture of a Decision Transformer is based on the original Transformer architecture,
with modifications to adapt it for reinforcement learning tasks. The key components of the
Decision Transformer architecture are:

15. 15/22
Input representation: The input for the Decision Transformer consists of state-action-
reward (SAR) tuples from the agent’s interaction with the environment. These tuples
are represented as a sequence of tokens, which are then embedded into continuous
vectors using an embedding layer.
Positional encoding: Positional encodings are added to the embedded input vectors
to capture the relative position of the input tokens. This allows the model to recognize
the order of the input tokens and understand the temporal relationships between them.
Multi-head self-attention: The core of the Transformer architecture is the multi-head
self-attention mechanism. This component computes attention scores for each input
token, capturing the dependencies between different tokens in the input sequence. The
self-attention mechanism is applied multiple times in parallel (multi-head) to learn
different aspects of the input data.
Feed-forward neural network: A feed-forward neural network is applied to each
position independently after the multi-head self-attention layer. This component
consists of two linear layers with an activation function (usually ReLU) in between,
which helps the model learn more complex patterns within the input data.
Residual connections and layer normalization: In each layer of the Transformer,
residual connections combine the outputs of the self-attention and feed-forward layers
with their inputs. Layer normalization is then applied to stabilize the training process
and improve the model’s generalization capability.
Stacking layers: The Decision Transformer’s architecture comprises multiple layers of
self-attention and feed-forward components stacked on top of each other. This deep
structure enables the model to learn complex hierarchical relationships within the input
data.
Output layer: The final layer of the Decision Transformer is a linear layer that maps
the continuous output vectors back to the action space, generating the predicted action
sequence.
Training objective: The Decision Transformer is trained to predict the action sequence
that maximizes the agent’s cumulative reward. This is achieved by minimizing the
cross-entropy loss between the predicted and ground truth action sequences derived
from the input SAR tuples.
In the Decision Transformer architecture, reinforcement learning is integrated implicitly
through the way the input data is represented and the training objective is defined. The
reinforcement learning aspects can be found in the following components:
Input representation: The input consists of state-action-reward (SAR) tuples collected
from the agent’s interaction with the environment. This data directly links the
architecture to the reinforcement learning problem, as it captures the agent’s
experience and the corresponding rewards.

16. 16/22
Training objective: The training objective in the Decision Transformer is designed to
predict the action sequence that maximizes the agent’s cumulative reward. By
minimizing the cross-entropy loss between the predicted action sequence and the
ground truth action sequence, the model learns to make decisions that result in higher
rewards.
While the architecture itself is built on the Transformer structure, the reinforcement learning
aspects are incorporated through the input representation and training objective. The
Decision Transformer does not explicitly include traditional reinforcement learning
components such as value functions or policy gradients. Instead, it leverages the power of
the Transformer architecture to learn an effective policy for the given task by predicting future
action sequences that maximize cumulative rewards.
How does the Decision Transformer work?
The Decision Transformer model works by leveraging the Transformer architecture to learn a
policy for reinforcement learning tasks from collected trajectory data. Here’s a detailed
explanation of its learning process:
Learning from collected trajectory data
The Decision Transformer uses offline reinforcement learning, meaning it works by learning
from previously collected data, which includes sequences of states, actions, and rewards
(called trajectories). These trajectories come from past experiences of other agents or even
human demonstrations.
To understand how the Decision Transformer works, imagine you have a dataset of these
trajectories representing different experiences and their corresponding rewards. The model
takes in the current situation and information about past experiences and tries to predict the
best sequence of actions that would lead to the highest total reward.
In simpler terms, the Decision Transformer looks at what has been done before in similar
situations and uses this knowledge to decide what actions to take in the current situation to
achieve the best outcome.
Handling long-term dependencies
One of the key challenges in reinforcement learning is handling long-term dependencies,
where the agent must make decisions based on information that may not be immediately
available or only becomes relevant after a series of steps. Transformers, by design, are well-
suited to address this issue due to their ability to capture contextual information and
dependencies across long sequences.

17. 17/22
The Decision Transformer benefits from the self-attention mechanism in Transformers, which
allows the model to focus selectively on different parts of the input sequence and learn long-
range dependencies between the states, actions, and rewards. As a result, the model can
capture the relationships between past experiences and future decisions, effectively learning
a policy that accounts for both immediate and long-term rewards.
By combining the powerful representation learning capabilities of Transformers with
trajectory data from reinforcement learning tasks, the Decision Transformer is able to learn
effective policies that consider both short-term and long-term dependencies. This results in
an agent capable of making decisions that maximize the cumulative reward over time,
addressing one of the major challenges in reinforcement learning.
Applications and use cases of the Decision Transformer
The Decision Transformer has a wide range of applications and use cases due to its ability to
learn from trajectory data and handle long-term dependencies. Some possible applications
and use cases include:
Robotics: Decision Transformers can be used to train robotic agents to perform
complex tasks, such as navigation, manipulation, and interaction with their
environment, by learning from collected demonstrations or past experiences.
Autonomous vehicles: In the context of self-driving cars, Decision Transformers can
learn optimal decision-making policies from collected driving data, enabling safer and
more efficient navigation in various traffic conditions.
Games: Decision Transformers can be employed to train agents that excel in strategic
games, like Go, chess, or poker, by learning from a dataset of expert gameplays and
mastering long-term planning.
Finance: Decision Transformers can be applied to optimize trading strategies, portfolio
management, and risk assessment by learning from historical financial data and taking
into account long-term dependencies.
Healthcare: In medical decision-making, Decision Transformers can be used to
optimize treatment plans, diagnose diseases, or predict patient outcomes by learning
from historical patient data and understanding the long-term effects of different
interventions.
Natural Language Processing (NLP): Although not a direct application of
reinforcement learning, Decision Transformers can be adapted to solve NLP tasks that
involve making decisions or generating responses based on long-term dependencies in
text data, such as machine translation, text summarization, or dialogue systems.
Supply chain optimization: Decision Transformers can be used to optimize inventory
management, demand forecasting, and routing decisions in complex supply chain
networks by learning from historical data and considering long-term effects.

18. 18/22
Energy management: Decision Transformers can help optimize energy consumption,
demand response, and grid stability in smart grid applications by learning from
historical usage data and considering long-term dependencies.
How to use the Decision Transformer in a Transformer?
Scenario: Here, we are demonstrating a Decision Transformer model, created by
Huggingface. This model is trained to become an ‘expert’ using offline reinforcement
learning. It’s specifically designed to operate effectively in the Gym Walker2d
environment.This model learns how to make optimal decisions for a two-legged walking
agent without directly interacting with the environment, using past experiences and data
collected from other sources. The model then applies its knowledge to help the agent
navigate and perform tasks effectively in the Gym Walker2d setting.
Step 1: Install the package
pip install git+https://github.com/huggingface/transformers
Step 2: Initiating the model
Employing the Decision Transformer is fairly straightforward; however, due to its
autoregressive nature, attention must be given to properly organizing the model’s inputs at
every time step. HuggineFace has provided a Python script and a Colab notebook to
illustrate the utilization of this model.
Use the below code to load pre-trained Decision Transformer model
from transformers import DecisionTransformerModel
model_name = "edbeeching/decision-transformer-gym-hopper-expert"
model = DecisionTransformerModel.from_pretrained(model_name)
Step 3: Establishing the environment
HuggingFace offers pre-trained checkpoints for Gym Hopper, Walker2D, and Halfcheetah
environments.
import gym
env = gym.make("Hopper-v3")
state_dim = env.observation_space.shape[0] # state size
act_dim = env.action_space.shape[0] # action size

19. 19/22
Step 4: Establishing autoregressive prediction function
The model executes an autoregressive prediction process, where predictions made at the
current time-step t rely on outputs from earlier time-steps. This function contains
considerable detail, which we will strive to elucidate in the comments.
# Function that gets an action from the model using autoregressive prediction
# with a window of the previous 20 timesteps.
def get_action(model, states, actions, rewards, returns_to_go, timesteps):
# This implementation does not condition on past rewards
states = states.reshape(1, -1, model.config.state_dim)
actions = actions.reshape(1, -1, model.config.act_dim)
returns_to_go = returns_to_go.reshape(1, -1, 1)
timesteps = timesteps.reshape(1, -1)
# The prediction is conditioned on up to 20 previous time-steps
states = states[:, -model.config.max_length :]
actions = actions[:, -model.config.max_length :]
returns_to_go = returns_to_go[:, -model.config.max_length :]
timesteps = timesteps[:, -model.config.max_length :]
# pad all tokens to sequence length, this is required if we process batches
padding = model.config.max_length - states.shape[1]
attention_mask = torch.cat([torch.zeros(padding), torch.ones(states.shape[1])])
attention_mask = attention_mask.to(dtype=torch.long).reshape(1, -1)
states = torch.cat([torch.zeros((1, padding, state_dim)), states], dim=1).float()
actions = torch.cat([torch.zeros((1, padding, act_dim)), actions], dim=1).float()
returns_to_go = torch.cat([torch.zeros((1, padding, 1)), returns_to_go], dim=1).float()
timesteps = torch.cat([torch.zeros((1, padding), dtype=torch.long), timesteps], dim=1)

20. 20/22
# perform the prediction
state_preds, action_preds, return_preds = model(
states=states,
actions=actions,
rewards=rewards,
returns_to_go=returns_to_go,
timesteps=timesteps,
attention_mask=attention_mask,
return_dict=False,)
return action_preds[0, -1]
Step 5: Model evaluation
To assess the model, we require further information, specifically the mean and standard
deviation of the states utilized during training. Luckily, this data is accessible for each
checkpoint’s model card on the Hugging Face Hub!
Additionally, we need a target return for the model. This highlights the strength of return-
conditioned offline reinforcement learning. By adjusting the target return, we can control the
performance of the policy effectively. This capability could be extremely valuable in a
multiplayer scenario, where we might want to fine-tune the difficulty of an opponent bot to
match the player’s skill level.
TARGET_RETURN = 3.6 # This was normalized during training
MAX_EPISODE_LENGTH = 1000
state_mean = np.array(
[1.3490015, -0.11208222, -0.5506444, -0.13188992, -0.00378754, 2.6071432,
0.02322114, -0.01626922, -0.06840388, -0.05183131, 0.04272673,])
state_std = np.array(
[0.15980862, 0.0446214, 0.14307782, 0.17629202, 0.5912333, 0.5899924,
1.5405099, 0.8152689, 2.0173461, 2.4107876, 5.8440027,])

21. 21/22
state_mean = torch.from_numpy(state_mean)
state_std = torch.from_numpy(state_std)
state = env.reset()
target_return = torch.tensor(TARGET_RETURN).float().reshape(1, 1)
states = torch.from_numpy(state).reshape(1, state_dim).float()
actions = torch.zeros((0, act_dim)).float()
rewards = torch.zeros(0).float()
timesteps = torch.tensor(0).reshape(1, 1).long()
# take steps in the environment
for t in range(max_ep_len):
# add zeros for actions as input for the current time-step
actions = torch.cat([actions, torch.zeros((1, act_dim))], dim=0)
rewards = torch.cat([rewards, torch.zeros(1)])
# predicting the action to take
action = get_action(model,
(states - state_mean) / state_std,
actions,
rewards,
target_return,
timesteps)
actions[-1] = action
action = action.detach().numpy()
# interact with the environment based on this action
state, reward, done, _ = env.step(action)
cur_state = torch.from_numpy(state).reshape(1, state_dim)

22. 22/22
states = torch.cat([states, cur_state], dim=0)
rewards[-1] = reward
pred_return = target_return[0, -1] - (reward / scale)
target_return = torch.cat([target_return, pred_return.reshape(1, 1)], dim=1)
timesteps = torch.cat([timesteps, torch.ones((1, 1)).long() * (t + 1)], dim=1)
if done:
break
The code block executes the interaction between the Decision Transformer model and the
environment, updating the model’s inputs and tensors as necessary while taking steps within
the environment.
Endnote
The Decision Transformer model transforms the way reinforcement learning agents are
trained and interact with their environments. By merging the best of both worlds – the
unparalleled ability of transformers to process sequential data and the dynamic learning
capabilities of reinforcement learning – Decision Transformers unlock new possibilities for a
range of applications and industries.
With their efficient offline learning process and proficiency in handling long-term
dependencies, Decision Transformers address the limitations of traditional RL approaches
and unlock new potential for creating advanced, intelligent systems. The applications for
Decision Transformers are vast and far-reaching, from autonomous vehicles and robotics to
strategic gameplay and personalized user experiences.
As we look to the future of AI, the Decision Transformer model stands as a testament to the
power of innovation and interdisciplinary research. By embracing the synergy between
Transformer architectures and reinforcement learning, we can unlock a new era of AI-
powered solutions, enhancing our capabilities and improving our understanding of the world
around us.
Leverage advanced Decision Transformer models to enhance AI-driven customer
experiences. Contact LeewayHertz’s Transformer development services for your consultancy
and development needs!