Learning Game-Theoretic Models of Multiagent Trajectories Using Implicit Layers

Authors: Philipp Geiger, Christoph-Nikolas Straehle4950-4958

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In experiments, we evaluate our approach on two real-world data sets, where we predict highway drivers merging trajectories, and on a simple decision-making transfer task.
Researcher Affiliation Industry Philipp Geiger, Christoph-Nikolas Straehle Bosch Center for Artificial Intelligence, Renningen, Germany philipp.w.geiger@de.bosch.com, christoph-nikolas.straehle@de.bosch.com
Pseudocode Yes (Its forward pass is explicitly sketched in Alg. 1 in Sec. B in (Geiger and Straehle 2021).) It contains the following modules Training of the architecture in principle happens as usual by fitting it to past scenes in the training set, sketched in Alg. 2 in Sec. B in (Geiger and Straehle 2021).
Open Source Code Yes Code is available at: https://github.com/boschresearch/trajectory_games_learning
Open Datasets Yes 1st data set: We use the high D data set (Krajewski et al. 2018), which consists of car trajectories recorded by drones over several highway sections. It is increasingly used for benchmarking (Rudenko et al. 2019; Zhang et al. 2020). From this data set, we use the recordings done over a section with an on-ramp. 2nd data set: We publish a new data set with this paper, termed HEE (Highway Eagle Eye). It consists of 12000 individual car trajectories ( 4h), recorded by drones over a highway section (length 600m) with an entry lane. The link to the data set and further details are in Sec. D.2 in (Geiger and Straehle 2021).
Dataset Splits Yes We use four-fold cross validation (splitting the data into 4 75% train and 25% validation).
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory, or cloud instance types) used for running experiments.
Software Dependencies No The paper does not list specific software dependencies with version numbers.
Experiment Setup Yes We use validation-based early stopping. We combine equilibrium refinement and weighting net into one module, consisting of two nets that predict the weights (kˆq)k K on the combination of (1) merging order (before/after) probabilities via a cross-entropy loss (2 hidden layers: 1 16, 1 4 neurons; dropout 0.6), and (2) Gaussian distribution over merging time point (discretized and truncated, thus the support inducing a refinement; 2 hidden layers: 1 64, 1 32 neurons; dropout 0.6), given x. For the preference revelation net we use a feed forward net (two hidden layers: 1 16, 1 24 neurons).22 As training loss we use mean absolute error (MAE; see also evaluation details below).