Paper here: https://iopscience.iop.org/article/10.1088/2632-2153/ad7571
AI planning using matrix product states to model sequences and the action selection algorithm provided by active inference. A model-based approach to learning and planning i.e. the agent’s generative model is a tensor network. Uses an update scheme which alternates between single-site and two-site updates, and allows dynamic variation of bond dimensions, while avoiding exploding bond dimensions as well as speeding up computation.
Two experiment files can be found inside mpstwo: frozenlake.py and tmaze.py.
E.g. to run frozenlake experiments with default hyperparameters:
python main.py frozenlake run
Hyperparameters can be adjusted through the command line:
python main.py frozenlake run --log_dir experiment1 --environment.is_slippery False
Alternatively, run experiments through:
python mpstwo/frozenlake.py run
Or with hyperparameters:
python mpstwo/frozenlake.py run model.init_mode="positive" model.cutoff=0.1 pool.rollouts=204
The hyperparameters which can be adjusted can be found in the config variable near the top of each file.