Tangled Program Graph (TPG) is a Reinforcement Learning (RL) technique based on genetic programming concepts. On state-of-the-art learning environments, TPGs have been shown to offer comparable competence with Deep Neural Networks (DNNs), for a fraction of their computational and storage cost.The contribution of this paper focuses on accelerating the inference of pre-trained TPGs, through the generation of standalone C code. While the training process of TPGs, based on genetic evolution principles, requires the use of flexible data structures supporting random mutations, this flexibility is no longer needed when focusing on the inference process.Evaluation of the proposed approach on four computing platforms, including embedded CPUs, produces an acceleration of the TPG inference by a factor 50 compared to state-ofthe-art implementations. The inference performance obtained within a complex RL environment range between hundreds of nano-seconds to micro-seconds, making this approach highly competitive for edge Artificial Intelligence (AI).

The integration of static parameters into Synchronous Dataflow (SDF) models enables the customization of an application functional and non-functional behaviours. However, these parameter values are generally set by the developer for a manual Design Space Exploration (DSE). Instead of a single value, moldable parameters accept a set of alternative values, representing all possible configurations of the application. The DSE is responsible for selecting the best parameter values to optimize a set of criteria such as latency, energy, or memory footprint. However, the DSE process explodes in complexity with the number of parameters and their possible values. In this paper, we study an automated DSE algorithm exploring multiple configurations of a dataflow application. Our experiments show that: 1) Only limited sets of configurations lead to Pareto-optimal solutions in a multi-criteria optimization scenario. 2) How individual parameters impact on optimization criteria are determined accurately from a limited subset of design points. The approach was evaluated on three image processing applications having from hundreds to thousands configurations.