Forem: Doraking

Re-reading Blockchain "Classics" in 2026: Implementing the Future in the Era of AI & Quantum Computing

Doraking — Thu, 29 Jan 2026 16:24:15 +0000

In the fast-moving Web3 industry, reading technical books published several years ago—now considered "classics"—in 2026 holds significant value.

Why? Because the "Why" (the fundamental philosophy) behind the technology remains far more vivid in early literature than in the latest documentation.

In this article, based on three classic books, I will revisit the foundational theories of blockchain, check the "Future Predictions" made back then against the "Current Solutions (2026 Implementations)," and explore where blockchain is heading now that AI and Quantum Computing have entered the practical stage.

📚 The "Bibles" Referenced

"Blockchain: Theory and Practice" (Similar scope to Mastering Bitcoin) (Japanese Book)
"The Textbook for Understanding Blockchain Structure and Development" (Japanese Book)
"Blockchain Revolution" by Don Tapscott and Alex Tapscott

Chapter 1: The Structure of "Trust" Redefined by Blockchain

Before diving into technical explanations, let's clarify the definition of "Trust" at the root of the architecture.

1.1 Mathematical Assurance of Social Capital

As stated in Blockchain Revolution, an ideal society is one rich in "Social Capital," where "courage worthy of praise is rewarded."

Traditionally, this "courage" (honest behavior) was guaranteed by Centralized Intermediaries (TTP: Trusted Third Parties) like banks and governments. However, these can become "Single Points of Failure (SPOF)" and incur mediation costs.

Blockchain replaced this trust with "Cryptographic Proof" and "Game Theoretic Incentives."

Cost of lying > Profit gained from lying

As long as this inequality holds, the system autonomously continues to record the "truth." This is the true meaning of "Trustless." It doesn't mean "trust no one"; it means "you trust the protocol so you don't have to trust anyone else."

1.2 Web3 and the "Internet of Value"

The concept proposed in early literature has become clearer today:

Web2 (Internet of Information): Copying is easy. Democratization of information.
Web3 (Internet of Value): Copying is impossible (prevention of double-spending). Democratization of value.

"Value" here isn't limited to currency. It includes identity, copyright, voting rights, and the "autonomous economic activities of AI" discussed later.

Chapter 2: Technical Details Supporting Bitcoin's Robustness

Let's look back at the mechanisms of Bitcoin with high resolution.

2.1 The Beauty of State Management: The UTXO Model

In Bitcoin, there is no variable called "Balance" in the database like in a bank account. There is only a collection of UTXOs (Unspent Transaction Outputs).

Input: References a past transaction (where the money came from).
Output: Specifies the new owner (who the rights are transferred to).
Script: A simple program (stack-based language) that verifies the validity of the transaction.

graph LR
    A[Tx A: 10 BTC] -->|Input| B(Tx B)
    B -->|Output 1| C[To Alice: 3 BTC]
    B -->|Output 2| D[To Bob: 7 BTC]
    style A fill:#f9f,stroke:#333,stroke-width:2px
    style B fill:#bbf,stroke:#333,stroke-width:2px
    style C fill:#bfb,stroke:#333,stroke-width:2px
    style D fill:#bfb,stroke:#333,stroke-width:2px

This model has the advantages of "ease of parallel processing" and "high privacy (easy to use disposable addresses)." On the other hand, it is unsuitable for complex state management (stateful processing) like Ethereum.

2.2 The Key to Tamper Detection: Merkle Trees

The block header contains the "Merkle Root," which is a summary value of all transactions included in that block.

Mechanism: A binary tree structure where transaction hashes are paired and hashed repeatedly until finally one hash value (the root) remains.
Benefit (SPV Nodes): Lightweight nodes (SPV) like smartphones can verify "whether a specific transaction is included in a block" using only the Merkle Path ($O(\log n)$ complexity) without downloading all data.

2.3 Consensus: The Essence of Proof of Work (PoW)

PoW is not just a calculation competition. It is a converter that transforms "physical world resources (electricity/hardware)" into "digital world security."

Nonce: Searching for a 32-bit integer such that the hash value is less than or equal to the Difficulty Target.
Probabilistic Finality: If you wait for "6 confirmations" (about 1 hour), the probability of the transaction being overturned becomes astronomically low.

Chapter 3: The 2026 "Answer Key" — Scalability and Ethereum

When these books were published (around 2017-2020), the biggest challenge was Scalability. What happened to the "prophecies" of that time?

3.1 The Defeat of Plasma and the Victory of Rollups

"Plasma," a scaling technology that was once considered promising in early technical texts, is almost unused in 2026.

Plasma's Failure: The Data Availability Problem. Because it only wrote the hash of the data to the parent chain, if the Operator maliciously hid the data, users risked being unable to withdraw funds.
Rollup's Victory: They adopted a method of writing "transaction data (compressed)" to L1 (Ethereum) as well as processing results. This allowed them to fully inherit L1 security.

Today, Optimistic Rollups (using fraud proofs) and zk-Rollups (using zero-knowledge proofs) are the center of the ecosystem.

3.2 Account Model and Smart Contracts

Ethereum adopted the Account Model instead of UTXO. This allowed it to hold "State" on the EVM (Ethereum Virtual Machine), enabling complex applications (DApps).

EOA (Externally Owned Account): Users holding private keys.
CA (Contract Account): The code itself.

The trend in 2026 is "Account Abstraction (ERC-4337)," which makes EOAs programmable like CAs. This is solving the biggest UX challenge of Web3: "Loss of Private Key = Loss of Assets."

Chapter 4: AI × Blockchain — The Arrival of the Agentic Economy

From here, based on the knowledge from the books, I will explain the most important topic of 2026: Fusion with AI.

4.1 "Bank Accounts" for AI Agents

The books talked about "IoT × Blockchain," but the true form was "AI Agent × Blockchain."
While it is a high hurdle for humans to operate Web3 wallets, for AI, Smart Contracts are the easiest APIs to handle.

Micropayments: AI pays usage fees for inference APIs in real-time, in units of $0.0001.
Autonomous Asset Management: AI agents manage and operate their own funds using DeFi protocols.

4.2 Proof of Humanity / Content Authenticity

With Generative AI flooding the world, technology to prove "whether this was made by a human or an AI" is essential.
Blockchain (public ledger) plays a vital role here as well. By carving the hash of the creation process into the chain, the Origin of digital content becomes immutable.

Chapter 5: Cryptography in the Quantum Era

Finally, I will touch on "Quantum Resistance," a topic I am particularly watching.
While classical texts state that ECDSA (Elliptic Curve Digital Signature Algorithm) is secure, if a quantum computer capable of implementing Shor's algorithm appears, the security of current blockchains will collapse.

5.1 Specific Threats

Deriving Private Key from Public Key: ECDSA and RSA rely on the difficulty of prime factorization or discrete logarithm problems, but quantum computing can solve these in polynomial time.
Hash Function Collisions: Grover's algorithm reduces the computational complexity of hash searching to $O(\sqrt{N})$ (although this is said to be manageable by doubling the key length).

5.2 The 2026 Countermeasure: Migration to PQC

Currently, major chains like Ethereum are preparing to migrate to Post-Quantum Cryptography (PQC).

Lattice-based cryptography: A leading candidate for quantum resistance.
STARKs: A Zero-Knowledge Proof technology considered quantum-resistant because it relies only on hash functions.

Precisely because blockchain promises to "preserve past records forever," cryptographic design anticipating computational power 10 or 20 years from now is required today.

Conclusion: Know the Theory, Implement the Future

The mathematical beauty told in Blockchain Theory.
The gritty implementation details learned in Textbooks.
The enthusiasm for social change depicted in Revolution.

These three perspectives have not faded even in 2026. Rather, with the addition of new variables like AI and Quantum Technology, their importance has increased.

What engineers need now is the ability to oscillate between the "Micro Perspective" (understanding EVM opcodes) and the "Macro Perspective" (AI economy and Quantum resistance).

The idea of "P2P Electronic Cash" presented by Satoshi Nakamoto in the white paper is now evolving into a "Foundation for an Economic Zone where Autonomous AI flies about." If you are going to participate in this grand experiment, now might be the most exciting time.

Quantum Error Correction Zootopia

Doraking — Wed, 24 Dec 2025 15:26:56 +0000

On the journey of learning quantum computing, there is a vast labyrinth that everyone wanders into at least once. Its name is "Quantum Error Correction (QEC)".

Opening textbooks, you encounter veterans like "Shor codes" and "Surface codes," but once you venture into the sea of research papers, you find countless "new species of codes" inhabiting it. The place that comprehensively collects and classifies such codes is the Error Correction Zoo (EC Zoo), managed by Victor V. Albert and others.

Looking at this site, a certain movie's worldview suddenly overlaps. Yes, it's Zootopia.

Imagine a world where diverse animals, from small mice to giant elephants, coexist under the common law called "Stabilizer Formalism." In this article, I will liken the residents of the EC Zoo to the characters of Zootopia and introduce them along with their mathematical definitions.

1. Topological Square: The Bond of Partners

The center of the story is the duo of topological codes active on the front lines of current quantum computer development (on 2D chips).

🐰 Judy Hopps = Surface Code

"Any hardware can implement me!"
The absolute protagonist of the current QEC world. Defined only by adjacent interactions on a 2D grid, it can be implemented on many hardware platforms such as superconducting circuits and neutral atoms.

【Overview】
The surface code is defined on an $\times L$ 2D square lattice. Physical qubits are placed on the edges (or vertices) of the lattice, and stabilizer operators are defined at vertices (Star) and faces (Plaquette).

A_v = \prod_{i \in \text{star}(v)} X_i, \quad B_p = \prod_{i \in \text{boundary}(p)} Z_i

All of these commute with each other ( $A_v, B_p] = 0$ ), and the code space $C\mathcal{C}$ is defined as the state having an eigenvalue of +1 for all stabilizers.

\mathcal{C} = { |\psi\rangle \mid A_v |\psi\rangle = |\psi\rangle, B_p |\psi\rangle = |\psi\rangle, \forall v, p }

Logical operators $XˉL,ZˉL\bar{X}_L, \bar{Z}_L$ are formed as strings crossing the lattice. This structure of "protecting the whole with only local checks" is exactly her straightforward investigative style.

Original Paper:
S. B. Bravyi and A. Y. Kitaev, "Quantum codes on a lattice with boundary" (1998)

🦊 Nick Wilde = Color Code

"You guys are simple black and white (X or Z), but I'm three colors."
A topological code like Judy, but with a more complex "trivalent graph (a lattice paintable with three colors)." He has the dexterity to perform specific logical gates like magic.

【Overview】
Color codes are typically defined on 2D hexagonal lattices (like honeycomb structures), where each face is painted with one of three colors (R, G, B).
Unlike the surface code, both $X$ -type and $Z$ -type stabilizers are defined for each face $f$ .

S_X^f = \prod_{i \in f} X_i, \quad S_Z^f = \prod_{i \in f} Z_i

Nick's greatest weapon (his con-artist dexterity) is the ability to implement transversal Clifford gates.
While implementing a Hadamard gate $H$ on a normal surface code requires complex operations, on a color code, a parallel operation $H⊗nH^{\otimes n}$ on each physical bit directly becomes a logical Hadamard $Hˉ\bar{H}$ .

\bar{H} = \bigotimes_{i=1}^n H_i

Original Paper:
H. Bombin and M. A. Martin-Delgado, "Topological Quantum Distillation" (2006)

2. Zootopia Police Department (ZPD): Organization and Discipline

The police department protecting the peace of the city has codes with unique but strict structures.

🐆 Benjamin Clawhauser = Steane Code

"I love donuts (holes) and CSS!"
The receptionist is the friendly $[[7, 1, 3]]$ code. A representative CSS code based on the classical Hamming code $[7, 4, 3]$ , it is a basic form loved by everyone.

【Overview】
The Steane code is a CSS code defined using a classical parity check matrix $H_{cl}$ .
For 7 physical bits, it has the following stabilizer generators:

\begin{aligned} g_1 &= IIIXXXX \\ g_2 &= IXXIIXX \\ g_3 &= XIXIXIX \\ g_4 &= IIIZZZZ \\ g_5 &= IZZIIZZ \\ g_6 &= ZIZIZIZ \end{aligned}

The first three are $X$ stabilizers, and the last three are $Z$ stabilizers. He is also used as a building block for Chief Bogo (mentioned later), serving as the beloved mascot and basic unit of the organization.

Original Paper:
A. M. Steane, "Error Correcting Codes in Quantum Theory" (1996)

🐃 Chief Bogo = Quantum Golay Code

"I will not tolerate any objections to my symmetry!"
A heavyweight with a massive and robust structure of [[23, 1, 7]]. A quantum version of the classical Golay code, boasting extremely high symmetry and defensive power.

【Overview】
The quantum Golay code is created from the classical "perfect code," the Golay code $G23\mathcal{G}_{23}$ , using the CSS construction method.
The code parameters are $[[23, 1, 7]]$ . That is, it uses 23 qubits and has a distance of $d = 7$ (can correct up to 3 errors).

Its stabilizer group $S\mathcal{S}$ is deeply related to sporadic simple groups like the Mathieu groups $M_{23}$ and $M_{24}$ , possessing mathematically extremely beautiful (and strict) symmetry.
While lacking the flexibility of topological codes, it is an old-fashioned tough code that never lets go of an error once caught.

Original Paper:
A. M. Steane, "Error Correcting Codes in Quantum Theory" (1996)

3. City Hall and the Underworld: Light and Shadow

The powerful figures moving this city. Their abilities (formulas) are powerful and distinctive.

🦁 Mayor Lionheart = Shor Code

"I built this paradise (QEC)."
The first quantum error-correcting code in history, $[[9, 1, 3]]$ . The great founder who proved that quantum error correction is possible.

【Overview】
The Shor code is created by concatenating a 3-qubit "bit-flip code" and a "phase-flip code." The logical states $∣0L⟩,∣1L⟩|0_L\rangle, |1_L\rangle$ are described as follows:

\begin{aligned} |0_L\rangle &= \frac{(|000\rangle + |111\rangle)(|000\rangle + |111\rangle)(|000\rangle + |111\rangle)}{2\sqrt{2}} \\ |1_L\rangle &= \frac{(|000\rangle - |111\rangle)(|000\rangle - |111\rangle)(|000\rangle - |111\rangle)}{2\sqrt{2}} \end{aligned}

This redundant structure became the first shield to protect quantum states from decoherence.

Original Paper:
P. W. Shor, "Scheme for reducing decoherence in quantum computer memory" (1995)

🐑 Assistant Mayor Bellwether = Bacon-Shor Code

"I look like just an assistant mayor, don't I? But I have 'gauges'."
At first glance, she looks like a normal stabilizer code, but she is a subsystem code that manipulates "gauge degrees of freedom." She has the face of a mastermind who rewrites the system behind the scenes.

【Overview】
The Bacon-Shor code has not only a normal stabilizer group $S\mathcal{S}$ but also a gauge group $G\mathcal{G}$ .
In an $\times L$ lattice, gauge operators act on adjacent two qubits.

G_{ij}^X = X_{i,j} X_{i,j+1}, \quad G_{ij}^Z = Z_{i,j} Z_{i+1,j}

Stabilizers are defined by products of these gauge operators (e.g., $S = G_{i,j}^X G_{i,j+1}^X$ ).
Her terror lies in the ability to apply gauge operators without destroying logical information. This is called "Gauge Fixing," allowing dynamic manipulation of the error correction procedure (or the ending of the story) depending on the situation.

Original Paper:
D. Bacon, "Operator quantum error-correcting subsystems for self-correcting quantum memories" (2006)

🐀 Mr. Big = 5-qubit Code

"Ice 'em." (Wasteful qubits, that is.)
[[5, 1, 3]]. The theoretical minimum size. The don of the underworld with perfect symmetry who allows absolutely no waste.

【Overview】
The smallest code capable of correcting a single qubit error. Its stabilizer generators have a beautiful structure created by cyclically shifting Pauli operators.

\begin{aligned} g_1 &= X Z Z X I \\ g_2 &= I X Z Z X \\ g_3 &= X I X Z Z \\ g_4 &= Z X I X Z \\ g_5 &= Z Z X I X \end{aligned}

Since this code is not a CSS code, $X$ and $Z$ are mixed. Though small, it possesses the strongest correction ability, truly a "small but mighty" existence.

Original Paper:
R. Laflamme et al., "Perfect Quantum Error Correcting Code" (1996)

4. Unique Citizens: The Pinnacle of Diversity

Zootopia also has unique codes adapted to special environments.

🦥 Flash = Floquet Code

"Wait... until... the cycle... ends..."
Incomprehensible in a static picture. A dynamic code defined together with "time".

【Overview】
Floquet codes do not have a fixed stabilizer group but are defined by an Instantaneous Stabilizer Group (ISG) $S(t)\mathcal{S}(t)$ where the measurement basis changes at each time step $t$ .
For example, on a honeycomb lattice, measurements are switched at each time step as follows:

\pmod 3 = 0: \text{Check } XX \text{ on horizontal edges} \\ t \pmod 3 = 1: \text{Check } YY \text{ on diagonal edges} \\ t \pmod 3 = 2: \text{Check } ZZ \text{ on vertical edges}

The reaction is slow (you have to wait 3 steps for information to gather), but by continuing observations, logical qubits dynamically emerge.

Original Paper:
M. B. Hastings and J. Haah, "Dynamically Generated Logical Qubits" (2021)

🦊 Finnick = GKP Code

"Thought I was a baby? I'm infinite-dimensional on the inside."
Looks like a single mode (small), but inside possesses an infinite-dimensional Hilbert space of Continuous Variables.

【Overview】
The Gottesman-Kitaev-Preskill (GKP) code is defined in the phase space of position operator $q^\hat{q}$ and momentum operator $p^\hat{p}$ .
Stabilizers are described by the following displacement operators $D(α)D(\alpha)$ :

Sq=ei2πq^,Sp=e−i2πp^ S_q = e^{i 2\sqrt{\pi} \hat{q}}, \quad S_p = e^{-i 2\sqrt{\pi} \hat{p}}

The logical state (codeword) is expressed as a superposition of "grid points" in phase space.

|\bar{0}\rangle \propto \sum_{n \in \mathbb{Z}} |q = 2n\sqrt{\pi}\rangle

A con artist of a different species (actually highly functional) with a structure fundamentally different from discrete qubits (other animals).

Original Paper:
D. Gottesman, A. Kitaev, and J. Preskill, "Encoding a qubit in an oscillator" (2001)

Conclusion: Future Star

🎤 Gazelle = qLDPC Codes

"Try Everything! (Break the limits!)"
The new era superstar breaking the trade-off between rate and distance.

【Overview】
Currently, the most attention is on Quantum Low-Density Parity-Check (qLDPC) codes.
The parity check matrix $H$ is a sparse matrix, and the weights of rows and columns are kept to a constant order.

[H_X, H_Z^T] = 0, \quad \text{wt}(row), \text{wt}(col) = O(1)

Especially recent discoveries like Bivariate Bicycle codes have amazing properties where both logical bit count $k$ and distance $d$ scale linearly with physical qubit count $n$ ( $\propto n$ ).
Researchers around the world are enthusiastic about her stage, which scales efficiently beyond the "limits of surface codes (area law)".

Original Paper:
S. Bravyi et al., "High-threshold and low-overhead fault-tolerant quantum memory" (2024)

The EC Zoo is updated daily, and new species are born.
By deciphering the "DNA" called mathematical formulas, the personality and survival strategy of each code become visible. I hope you will also find your favorite code at the EC Zoo.

Theory and Implementation of AlphaQuoridor

Doraking — Thu, 01 Aug 2024 15:10:39 +0000

In Japan, the board game Quoridor is not well-known. This article applies AlphaZero, a powerful deep reinforcement learning method, to Quoridor. Despite its simple design, AlphaZero has demonstrated the ability to outperform professional players in games like Go and Shogi. The goal is to deepen understanding of both the theoretical and practical aspects of AlphaZero through this application.

What is Quoridor?

According to Wikipedia, Quoridor is a French abstract board game where players aim to reach the opposite side of the board by moving their pieces and placing walls.

Quoirdor — Wikipedia

The setup and rules are as follows:

Setup:

For two players, each starts at opposite ends of the board and has ten walls. For four players, each starts in their own corner with five walls.

Gameplay:

Players take turns either moving their piece one square in any direction or placing a wall on the board. Pieces can jump over other pieces but cannot cross walls. Walls cannot completely block a player’s path.

Winning:

The first player to reach the opposite side wins.
The simple rule that walls cannot completely block paths makes the game exciting until the end.

The simple rule that walls cannot completely block paths makes the game exciting until the end.

Theory of AlphaZero

AlphaZero [1, 2] combines deep learning, search, and reinforcement learning. Let’s explore each component:

Deep Learning:

AlphaZero uses deep learning for intuition, similar to how professional players think about their best moves. It employs ResNet [3], a convolutional neural network used in image analysis. The game board, like an image, is a 2D array of information. The network can be represented as:

ResNet(s)=π(a∣s),v(s)\mathrm{ResNet}(s) = \pi(a | s), v(s)

where $s$ 　is the game state, $π (a ∣ s)$ is the policy (probability of action a given state $s$ ), and $v (s)$ is the value (+1 if win,-1 if loss) of state $s$ .

Search:

Games like Shogi, Go, Othello, and Quoridor are two-player zero-sum games with perfect information. The optimal strategy can be found using the minimax method, but it is impractical due to the large number of possible moves. Instead, AlphaZero uses Monte Carlo Tree Search (MCTS) to predict future moves.

MCTS decides the next move based on the Upper Confidence Bound (UCB) value:

UCB(s,a)=Q(s,a)+c⋅π(a∣s)⋅N(s)1+N(s,a)\mathrm{UCB}(s, a) = Q(s,a) + c \cdot \pi(a | s) \cdot \frac{\sqrt{N(s)}}{1+N(s, a)}

\frac{1}{N(s, a)} \sum_{i=1}^{N(s, a)} v_i(s \stackrel{a}{\to} s^\prime)

where $Q (s, a)$ is the value estimate of action $a$ in state $s$ , averaged over the number of simulations $N (s, a)$ transitioning from state $s$ to $s′s^\prime$ . $N (s)$ is the total number of simulations for state $s$ . The first term of the equation emphasizes exploiting high-value actions, while the second term focuses on exploring actions that haven't been simulated much, balanced by the hyperparameter $c$ .

By repeating simulations, actions with higher $N (s, a)$ become valuable moves.

In practice, the action $a$ to be taken in a given state $s$ is determined by the probability $π (a ∣ s)$ , weighted by the Boltzmann distribution with temperature $T$ .

π(a∣s)=N(s,a)1/T∑bN(s,b)1/T\pi(a | s) = \frac{\displaystyle N(s, a)^{1/T}}{\displaystyle \sum_b N(s, b)^{1/T}}

Reinforcement Learning:

AlphaZero generates training data through self-play, updating ResNet parameters with this experience to create a more intelligent neural network.

The process involves:

Initializing ResNet parameters.
Obtaining experience data through self-play.
Updating ResNet parameters with the experience data.
Repeating steps 2 and 3 multiple times.

Implementation of AlphaQuoridor

The implementation of AlphaQuoridor involves several steps, with all code available on GitHub:

https://github.com/dorakingx/AlphaQuoridor

Game design (game.py)
Deep learning implementation with ResNet (dual_network.py)
Monte Carlo Tree Search implementation (pv_mcts.py)
Data collection through self-play (self_play.py)
Updating ResNet parameters (train_network.py)
Comparing and updating the best parameters (evaluate_network.py)
Evaluating the best player (evaluate_best_player.py)
Running the entire training cycle (train_cycle.py)
Implementing a game UI to play against the AI (human_play.py)

This project references the Japanese book "AlphaZero: Deep Learning, Reinforcement Learning, and Search" for detailed explanations.

AlphaZero: Deep Learning, Reinforcement Learning, and Search Practical Introduction to Artificial Intelligence Programming

Due to the short training time, the game was trained on a 3x3 board. While this isn't very exciting gameplay-wise, the implementation allows for easy adjustment to larger board sizes, such as the actual 9x9 size, which would demonstrate the reinforcement learning capabilities better.

The actual gameplay screen when running human_play.py looks like this. The number of walls for both the player and the enemy (AI) is set to one, matching the game size. To place a wall, click "Place Wall" below, select the vertical or horizontal direction, and then click the desired location (red grid points).

Initial state

Mid-game

End-game case for winning

End-game case for losing

The AI wasn't very strong this time due to the limited training time. However, with a longer training period, a stronger AI could be developed.

Summary

This article applied AlphaZero to the lesser-known Quoridor board game. AlphaZero can be adapted to any two-player zero-sum game with perfect information by modifying the game design. This could be an interesting project to see how strong an AI can be created for different games. If you enjoyed this article, please like it, and stay tuned for more articles.

References

[1] D. Silver et al., "Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm"
[2] D. Silver et al., "A General Reinforcement Learning Algorithm That Masters Chess, Shogi, and Go through Self-Play"
[3] K. He, X. Zhang, S. Ren, and J. Sun, "Deep Residual Learning for Image Recognition"