TASK Quarterly

Generative methods in classification tasks

2025-06-17T12:34:10+02:00

The paper presents the implementation of generative methods in classification tasks. A distinction is made between
two types of tasks – supervised learning and unsupervised learning – along with example use cases. Within the scope of
supervised methods, described are the Bayes classifier and the use of the multivariate Gaussian distribution. To solve
the unsupervised learning task using a generative approach, the Gaussian Mixture Model (GMM) is presented. The
paper also describes a generative neural network based on an autoencoder architecture, implemented as a Variational
Autoencoder (VAE).

Assessing AMBER and UNRES Force Fields on the Energy Landscape of Bovine Pancreatic Trypsin Inhibitor

2025-11-24T10:06:30+01:00

Biological complexity emerges from the collective behaviour of proteins, whose folding and conformational dynamics are directly governed by the underlying potential energy surface. Here, we compare the potential energy landscapes of bovine pancreatic trypsin inhibitor (BPTI) obtained with the all-atom AMBER force field and the coarse-grained UNRES potential. For the native triply disulphide-bonded state of BPTI, both models sample predominantly folded, native-like conformations with relatively small structural deviations from the crystallographic structure, though UNRES explores structures with a broader range of RMSDs compared to experimental structure, radii of gyration, and solvent-accessible surface areas. Using comparable CPU time, UNRES generates a more diverse set of minima and transition states, yielding a more extensively sampled landscape with a globally similar topology. Together, these results highlight how all-atom and coarse-grained potentials differ in the representation of protein energy landscapes, while retaining consistent global features for a highly constrained, disulphide-rich protein such as BPTI.

Interpreting Deep Q-Networks: A Rule-Based Comparison with First-Order Logic in Wumpus World

2025-08-05T12:41:35+02:00

Deep reinforcement learning models such as Deep Q-Networks (DQNs) have achieved great performance in both simple and complex environments, but their decision-making process remains largely opaque. This work addresses the interpretability challenge by proposing a~method of extracting and comparing symbolic rules from a trained DQN and logic-based agents. The method is showcased in the popular Wumpus World domain. Rules extraction from the agents is done via training decision trees to mimic the agent's behavior. Comparison is done using Jaccard similarity and simple structural metrics. Results show that despite similar performance, DQNs and logic agents rely on partially overlapping but structurally largely distinct decision rules. This highlights the feasibility of translating subsymbolic policies into interpretable rules and reveals meaningful structural differences between learned and symbolic strategies.