diff options
| -rw-r--r-- | doc/diagrams/planificationWorkPlanEngine.puml | 46 | ||||
| -rw-r--r-- | doc/diagrams/planificationWorkPlanGame.puml | 30 | ||||
| -rw-r--r-- | doc/tex/planification.tex | 197 | ||||
| -rw-r--r-- | imago/engine/keras/initialDenseNeuralNetwork.py | 28 |
4 files changed, 301 insertions, 0 deletions
diff --git a/doc/diagrams/planificationWorkPlanEngine.puml b/doc/diagrams/planificationWorkPlanEngine.puml new file mode 100644 index 0000000..9caad40 --- /dev/null +++ b/doc/diagrams/planificationWorkPlanEngine.puml @@ -0,0 +1,46 @@ +@startgantt + +!include skinparams.puml +!include skinparamsGantt.puml + +printscale weekly +Saturday are closed +Sunday are closed + +Project starts 2021-01-11 + +-- Preliminary research -- +[Previous works research] as [PWR] lasts 1 week +[Algorithms research] as [AR] lasts 2 weeks + +-- Engine Implementation -- +[Engine modelling] as [EM] lasts 1 week +[Engine implementation] as [EI] lasts 4 weeks + +-- Algorithms Implementations -- +[Monte Carlo implementation] as [MCI] lasts 4 weeks +[Neural networks research] as [NNR] lasts 2 weeks +[Neural networks implementation] as [NNI] lasts 3 weeks + +-- Testing -- +[Engine unit testing] as [EUT] lasts 4 weeks +[System testing] as [ST] lasts 1 week + +-- Analysis -- +[Algorithms comparison] as [AC] lasts 2 weeks + +[PWR] -> [AR] +[AR] -> [EM] + +[EM] -> [MCI] +[EM] -> [EI] +[EM] -> [EUT] + +[MCI] -> [NNR] +[NNR] -> [NNI] + +[NNI] -> [ST] + +[ST] -> [AC] + +@endgantt diff --git a/doc/diagrams/planificationWorkPlanGame.puml b/doc/diagrams/planificationWorkPlanGame.puml new file mode 100644 index 0000000..ffaf72c --- /dev/null +++ b/doc/diagrams/planificationWorkPlanGame.puml @@ -0,0 +1,30 @@ +@startgantt + +!include skinparams.puml +!include skinparamsGantt.puml + +printscale weekly zoom 2 +Saturday are closed +Sunday are closed + +Project starts 2020-11-02 + +-- Preliminary research -- +[Previous works research] as [PWR] lasts 1 week + +-- Game Implementation -- +[Domain modelling] as [DM] lasts 1 week +[Domain implementation] as [DI] lasts 6 weeks + +-- Testing -- +[Unit testing] as [UT] lasts 6 weeks +[System testing] as [ST] lasts 1 week + +[PWR] -> [DM] + +[DM] -> [DI] +[DM] -> [UT] + +[DI] -> [ST] + +@endgantt diff --git a/doc/tex/planification.tex b/doc/tex/planification.tex new file mode 100644 index 0000000..f0f7535 --- /dev/null +++ b/doc/tex/planification.tex @@ -0,0 +1,197 @@ +\section{Planning} + +\subsection{Driving Needs} + +As one of the deepest and most studied games in the world, Go presents a very +interesting problem for artificial intelligence. Implementing not only the +game's simple but subtle rules, but a system capable of playing it with a +satisfying level of skill, is a task worth of pursuing as an exercise on +software design, algorithmics and AI research. + +On the practical level, this project can be a foundation for the development of +different Go analysis algorithms by providing an existing engine to house them, +which can be of interest to Go players and software scientists alike. + +\subsection{Reach} + +Presented here are the ideal targets of the project. + +\begin{itemize} + \item An implementation of the game of Go, that is, a system for holding the + moves and variants of a match (a tree of moves) and the logic for the + game's rules. + \item An engine capable of analyzing board positions and generating strong + moves via various decision algorithms. + \item Either a GUI specifically developed for the project or an + implementation of an existing protocol so the engine can be used with + existing tools and GUIs. + \item A way for processing existing records of games, which are usually + recorded in the SGF format. +\end{itemize} + +\subsection{Project Stages} + +The project will be organized in several stages based on the different +components and needs. + +\subsubsection{Game Implementation} + +The rules of the game must be implemented, ideally in a way they can be tested +by direct human play. This system will at its bare minimum represent the +Japanese Go rules (area scoring, no superko rule, no suicide moves). + +\subsubsection{Engine Implementation} + +The key of this project is to create some kind of system able to generate strong +moves based on any given board configuration: this will be such system. It will +implement an existing protocol so it can be used with other compatible tools. It +has to be able to receive game updates and configuration and to output moves for +either player. It should also be modular enough so different algorithms can be +selected and tested against each other as an experimental search for the best of +them. + +\subsubsection{Artificial Intelligence Algorithms} + +Different algorithms for the engine to use should be implemented and tested. The +results of this development and testing process should be presented as part of +the final version of the project. + +\subsection{Logistics} + +The project will be developed by Íñigo Gutiérrez Fernández, student of the +Computer Software Engineering Degree at the University of Oviedo, with +supervision from Vicente García Díaz, Associate Professor in the Department of +Computer Science at the University of Oviedo. + +The used material consists of a development and testing machine owned by the +student with specifications stated later on the project plan. + +\subsection{Work Plan} + +The sole developer will be the student, who is currently working as a Junior +Software Engineer on a 35 hour per week schedule and with no university +responsibilities other than this project. Taking this into account, a sensible +initial assumption is that he will be able to work 3 hours a day, Monday to +Friday. Gantt diagrams for the planned working schedule are shown as +\fref{fig:planificationWorkPlanGame} and +\fref{fig:planificationWorkPlanEngine}. + +\begin{figure}[h] + \begin{center} + \includegraphics[width=\textwidth]{diagrams/planificationWorkPlanGame.png} + \caption{Initial work plan for implementing the game. + }\label{fig:planificationWorkPlanGame} + \end{center} +\end{figure} + +\begin{figure}[h] + \begin{center} + \includegraphics[width=\textwidth]{diagrams/planificationWorkPlanEngine.png} + \caption{Initial work plan for implementing the engine and algorithms. + }\label{fig:planificationWorkPlanEngine} + \end{center} +\end{figure} + +\subsection{Previous Works} + +\subsubsection{Existing Engines} + +\paragraph{AlphaGo} + +A Go play and analysis engine developed by DeepMind Technologies, a company +owned by Google. It revolutionized the world of Go in 2015 and 2016 when it +respectively became the first AI to win against a professional Go player and +then won against Lee Sedol, a Korean player of the highest professional rank and +one of the strongest players in the world at the time. Its source code is +closed, but a paper\cite{natureAlphaGo2016} written by the team and +published on Nature is available on +https://storage.googleapis.com/deepmind-media/alphago/AlphaGoNaturePaper.pdf. + +The unprecedented success of AlphaGo served as inspiration for many AI projects, +including this one. + +\paragraph{KataGo\cite{katago}} + +An open source project based on the AlphaGo paper that also achieved superhuman +strength of play. The availability of its implementation and documentation +presents a great resource for this project. + +\paragraph{GnuGo\cite{gnugo}} + +A software capable of playing Go part of the GNU project. Although not a strong +engine anymore, it is interesting for historic reasons as the free software +engine for which the GTP protocol was first defined. + +\subsubsection{Existing Standards} + +\paragraph{GTP\cite{gtp}} + +GTP (\textit{Go Text Protocol}) is a text based protocol for +communication with computer go programs. It is the protocol used by GNU Go and +the more modern and powerful KataGo. By supporting GTP the engine developed for +this project can be used with existing GUIs and other programs, making it easier +to use it with the tools users are already familiar with. + +\paragraph{SGF\cite{sgf}} + +SGF (\textit{Smart Go Format} or, in a more general context, \textit{Smart Game +Format}) is a text format widely used for storing records of Go matches which +allows for variants, comments and other metadata. It was devised for Go but it +supports other games with similar turn-based structure. Many popular playing +tools use it. By supporting SGF vast existing collections of games, such as +those played on online Go servers, can be used to train the decision algorithms +based on neural networks. + +\subsubsection{Sabaki\cite{sabaki}} + +Sabaki is a go board software compatible with GTP engines. It can serve as a GUI +for the engine developed in this project and as an example of the advantages of +following a standardized protocol. + +\subsubsection{Keras\cite{keras}} + +Keras is a deep learning API for Python allowing for the high-level definition +of neural networks. This permits easily testing and comparing different network +layouts. + +\subsection{Technological Infrastructure} + +\subsubsection{Programming Language}\label{sec:programmingLanguage} + +The resulting product of this project will be one or more pieces of software +able to be run locally on a personal computer. The programming language of +choice is Python\cite{python}, for various reasons: + +\begin{itemize} + + \item It has established a reputation on scientific fields and more + specifically on AI research and development. + \item Interpreters are available for many platforms, which allows the most + people possible to access the product. + \item Although not very deeply, it has been used by the developer student + during its degree including in AI and game theory contexts. + +\end{itemize} + +\subsubsection{Interface} + +Both the game and the engine will offer a text interface. For the game this +allows for quick human testing. For the engine it is mandated by the protocol, +since GTP is a text based protocol for programs using text interfaces. +Independent programs compatible with this interface can be used as a GUI. + +There is also the need of an interface with SGF files so existing games can be +processed by the trainer. + +Both the engine and the trainer will need to interface with the files storing +the neural network models. + +The systems' interfaces are shown in \fref{fig:interfaces}. + +\begin{figure}[h] + \begin{center} + \includegraphics[width=\textwidth]{diagrams/interfaces.png} + \caption{Interfaces of the three components of the project.} + \label{fig:interfaces} + \end{center} +\end{figure} diff --git a/imago/engine/keras/initialDenseNeuralNetwork.py b/imago/engine/keras/initialDenseNeuralNetwork.py new file mode 100644 index 0000000..dfe8379 --- /dev/null +++ b/imago/engine/keras/initialDenseNeuralNetwork.py @@ -0,0 +1,28 @@ +"""Dense neural network.""" + +from tensorflow.keras.models import Sequential +from tensorflow.keras.layers import Dense +from tensorflow.keras.optimizers import Adam + +from imago.engine.keras.neuralNetwork import NeuralNetwork + +defaultModelFile = 'models/imagoDenseKerasModel.h5' + +class DenseNeuralNetwork(NeuralNetwork): + + def _initModel(self, boardSize=NeuralNetwork.DEF_BOARD_SIZE): + model = Sequential([ + Dense(units=16, activation='relu', input_shape=(boardSize,boardSize)), + Dense(units=32, activation='relu'), + Dense(units=boardSize, activation='softmax') + ]) + + model.summary() + + model.compile( + optimizer=Adam(learning_rate=0.0001), + loss='categorical_crossentropy', + metrics=['accuracy'] + ) + + return model |