Modelling and explaining the basis of intelligent behaviour is a crucial problem today, given the exponential growth of artificial intelligence systems around us. Active inference, now regarded as a general theory of behaviour, offers a principle approach to probe the basis of intelligent behaviour. We summarise, compare, and test two major active inference schemes based on planning and learning from experience. Observing a data-complexity trade-off between the two schemes, we propose a mixed model that enables balanced decision-making. We test the proposed model in a complex grid-world environment that demands the agent's adaptability. We also note that our model allows us to study the evolution of parameters and reveal meaningful insights helping us to investigate the basis of intelligent decision-making.
aswinpaul/aimmppcl_2023
An Active Inference Mixed Model Integrating Predictive Planning and Counterfactual Learning
Jupyter NotebookMIT