This repository is a full AI/ML journey implemented across six projects, from classical search and evolutionary computation to supervised learning, unsupervised learning, and deep generative models.
The structure below intentionally uses topic names (instead of generic tp0, tp1, ..., tpN) so visitors can immediately see the real scope of the work.
| Topic Name | Folder | What is inside |
|---|---|---|
| Pokemon Capture Probability and Data Analysis | tp0 |
Data analysis workflow for modeling Pokemon capture function behavior |
| State-Space Search: 8-Puzzle + Sokoban | tp1 |
BFS, DFS, A*, Local Greedy, Global Greedy, heuristics, benchmarking |
| Genetic Algorithms for RPG Build Optimization | tp2 |
Evolutionary operators, adaptive mutation, selection/replacement strategies |
| Supervised Learning: Perceptrons and MLPs | tp3 |
Step/linear/non-linear perceptrons, MLPs, optimizer studies, noise robustness |
| Unsupervised Learning and Associative Memory | tp4 |
Kohonen SOM, PCA, Oja, Sanger, Hopfield |
| Autoencoders and Variational Latent Spaces | tp5 |
AE, DAE, VAE, latent-space exploration, custom image experiments |
- It covers both symbolic AI and statistical learning.
- It includes from-scratch implementations, not just high-level library wrappers.
- It focuses on experimentation: configuration-driven runs, metrics, plots, and visual analysis.
- It demonstrates breadth and depth across core AI foundations and modern ML methods.
- Search and planning: backtracking, repeated-state handling, informed vs uninformed search.
- Evolutionary computation: crossover/mutation design, dynamic mutation schedules, mixed selection policies.
- Supervised learning: perceptrons, MLP topology tuning, optimizer effects, noisy data robustness.
- Unsupervised learning: topology-preserving maps, dimensionality reduction, Hebbian rules.
- Associative memory: Hopfield pattern storage and recovery dynamics.
- Generative modeling: AE/DAE/VAE and latent-space geometry.
This project introduces the portfolio with a data analysis workflow focused on Pokemon capture probability modeling. It establishes reproducible experimentation through config-based execution and report-driven interpretation.
- Folder README:
tp0/README.md - Docs:
tp0/docs/SIA_TP0.pdf,tp0/docs/Informe-TP0.pdf - Run:
cd tp0
pipenv install
pipenv run python main.py ./configsThis project explores classical AI search deeply, combining algorithmic implementations with visual and metric-based comparisons.
- Implemented search methods include BFS, DFS, A*, Local Greedy, and Global Greedy.
- Includes heuristic experimentation and analysis pipelines for runtime, expanded nodes, and movement cost.
- Provides both interactive Sokoban visualization and reproducible batch experiments.
Folder docs:
tp1/README.md- Core algorithms:
tp1/core/algorithms - Heuristics:
tp1/core/heuristics.py
Representative visuals:
This project builds an optimization engine to find high-performing RPG character configurations under game-time constraints.
- Rich operator design: multiple crossover types, mutation modes, and mutation distributions.
- Dynamic mutation rates (constant, sinusoidal, exponential decay).
- Weighted composition of parent selection and replacement strategies.
- Extensive config system for reproducibility and hyperparameter analysis.
Folder docs:
tp2/README.md- Main run entry:
tp2/src/master.py
Representative visuals:
This project progresses from simple perceptrons to deeper MLP architectures, including practical tooling for training analysis and robustness testing.
- Step, linear, and non-linear perceptrons.
- MLP training for XOR, parity, and MNIST-like digit tasks.
- Optimizer comparisons (gradient descent, momentum, Adam).
- Noise-injection experiments and confusion-matrix evolution.
Folder docs:
tp3/README.md- Model implementations:
tp3/src/models
Representative visuals:
This project focuses on structure discovery and memory dynamics via classic unsupervised and recurrent neural approaches.
- Kohonen SOM for topology-preserving projection and clustering.
- PCA implementation via covariance analysis.
- Oja and Sanger rules for principal component extraction.
- Hopfield network for pattern storage and denoising-style recovery.
Folder docs:
tp4/README.md- Core implementations:
tp4/src/core
Representative visuals:
This project extends the neural stack into representation learning and generation with AE, DAE, and VAE.
- AE/DAE abstractions built over MLP foundations.
- VAE implementation with encoder/decoder pipelines and latent sampling.
- Latent-space trajectory visualizations for interpretability.
- Custom image workflow for face dataset experimentation.
Folder docs:
tp5/README.md- Core implementations:
tp5/src/core
Representative visuals:
All visual assets found in the repository are embedded below so the full experimentation trail is visible in one place.
Pokemon Capture Probability and Data Analysis (tp0, 0 assets)
No images were found in this folder.
State-Space Search: 8-Puzzle + Sokoban (tp1, 43 assets)
Genetic Algorithms for RPG Build Optimization (tp2, 19 assets)
Supervised Learning: Perceptrons and MLPs (tp3, 38 assets)
Unsupervised Learning and Associative Memory (tp4, 5 assets)
Autoencoders and Variational Latent Spaces (tp5, 9 assets)
Most folders use independent configs and environments. A practical flow is:
# example pattern
cd tpX
pipenv install
# run script defined in that folder READMEThis repository is designed as an AI/ML learning and engineering portfolio: each project focuses on a different problem class, and together they show end-to-end understanding of core concepts across search, optimization, supervised learning, unsupervised learning, and generative modeling.