Ryan Zhang
_
Full-Stack Developer with a passion for software and hardware
Projects
Click on my projects to see what I learnt during development!
Visit my Github for codebases & technical documentation
-
ProcessingCiv Game Development + Usage of LLMs in CPU AI
-
Foreign Travel Office Map Web Development + Web Scraping
-
FPGA Experiments FPGA Development + Computer Architecture
-
Maze Algorithms Graph Algorithms + App Design
-
ryan-zhang.com Web Development
ProcessingCiv
Overview
A simplified, in-progress, solo-developed port of Civilization 5 written fully in Processing Java, built as a playground for experiments with LLM-based CPU AIs.
What I learnt
This is the largest-scale project I worked on, with 5k+ lines written, and a wide scope ranging from terrain generation to local multiplayer.
The development process has taken ~2 months to reach ~70% completion of my project brief, created at the start.
Taught me: Performance optimization, improved my debugging skills, and made my coding practices more maintainable.
Movement
Legal moves, calculated by weighted BFS, have a blue boundary. Turn-terminating moves are marked with X.
Attack
In-range enemies have a red boundary. Melee attacks first move to an adjacent tile. Damage and counter-damage is calculated from the unit's strength, positioning and bonuses.
Capture
A melee unit can capture an enemy unit if it has 0 health (a passive unit, e.g. trader), or will drop to 0 health upon attack. Smaller icons choose who's in front when multiple units are stacked.
City Founding
A settler can found a city, on habitable, isolated land. A city can produce population, contributing to production, and several units and buildings, some of which are specific to the tile type the city is founded on.
New Unit
Newly-produced units can stack on the city tile, and are uncovered by clicking on the city header.
Ranged Attack
A ranged unit (e.g. archer) can attack all non-obstructed enemies within range. Arrows can't bend or shoot over hills/mountains.
Embarkation
A land unit may embark on water at a movement penalty and the ending of its turn. Valid embarkation tiles have dark blue boundaries.
City Sieging
A city can be attacked by melee and ranged units. Once it is attacked by a melee unit and it's health falls below 0, the city is annexed to the civilization of the melee unit, and its population erased.
City Bombardment
A city can bombard all enemy units in range not obstructed by mountains.
Disembarkation
A land/naval unit may disembark to land. Naval units receive a severe movement penalty unless it travels along a river tile.
Hardest Challenges During Development
Map Generation - Creating a realistic distribution of biomes and river generation
Using perlin noise for humidity, altitude and temperature noise maps, and adapting tree growth algorithms for pruning/merging rivers.
Local Multiplayer - Which game states to pass?
Use delta compression, game states that affect other players immediately (e.g. coming within range/attacking enemy unit) are transmitted to the player via host. Discrepancies are resolved each turn by host, residual data distributed.
Optimization - Rendering large maps is laggy...
Compatibilizing program with the P2D OpenGL renderer to utilize GPU hardware acceleration, implementing culling and level-of-depth for rendering, and quadtrees for clicking.
The LLM CPU
Overview
An LLM fed game states that performs analysis and considers real-world cases, outputting next-move decisions and long-term game plans. Inspired by the paper Richelieu: Self-Evolving LLM-Based Agents for AI Diplomacy.
Architecture
- Generic, important game states are translated into strictly-formatted text. An LLM agent tells us whether the current long-term plan is worth deviating or modifying based on updates to the previously-considered game state.
- An LLM agent ranks the importance of game systems (e.g. 1. Military Development, 2. Science Production)
- A set of high-level, prioritised instructions are made, e.g. "1. Continue invasion of X, 2. Research towards science-producing buildings". These are recorded for consideration in future turns.
- An LLM agent exists for each game system, with access to more specific game states for its focus area. These translate high-level instructions to low level, strictly formatted instructions. An LLM agent is used to assign tasks, based on priority and sequencing.
- In hardcode, low-level instructions are checked for conflicts, and parsed and executed by the game.
Trello Graph
Trello was used to organise and schedule development milestones, and ensure basic features were implemented before complex ones.
Graph Depicting LLM-based AI Architecture
Demonstrating the implementation of the multi-agent LLM in coming to a decision.
Graph Depicting Decision Tree for the Multi-Move Function
Used to decide whether to go through tracked, unfogged land using A* or track through a foggy area with greedy BFS.
Current State
Developed prototype for a military LLM agent. Made revisions in greater architecture to account for latency, context window saturation, semantic decay and coherence drift. Llama 70-B has been used during demos for its price, RAG performance and 128k context window.
ryan-zhang.com
Overview
A portfolio website showing off my projects.
What I learnt
This project took 5 days to make, refreshed my knowledge on SVG animation, CSS and Cloudflare web hosting.
Experimented with a minimalist-theme for the website, allowing more focus to be placed on the projects.
What I found most difficult: Words drifting off the avatar, color coordination, web responsiveness between desktop/laptop