Multi-Agent Intention Recognition and Progression

Authors: Michael Dann, Yuan Yao, Natasha Alechina, Brian Logan, Felipe Meneguzzi, John Thangarajah

IJCAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our approach in the two-agent version of the well-known Craft World [Andreas et al., 2017] environment developed by Dann et al. (2022) to evaluate IRM. and The experiment results are summarised in Tables 3, 4 and 5. All results are averaged over 500 randomly generated task instances, with the best results highlighted in bold.
Researcher Affiliation Academia 1RMIT University 2University of Nottingham, Ningbo China 3Utrecht University 4 University of Aberdeen
Pseudocode Yes Algorithm 1 Rollout phase for IGR (one other agent).
Open Source Code Yes Code is available at https://github.com/mchldann/IJCAI GR.
Open Datasets Yes We evaluate our approach in the two-agent version of the well-known Craft World [Andreas et al., 2017] environment developed by Dann et al. (2022) to evaluate IRM.
Dataset Splits No The paper mentions that Q-functions are trained across randomly generated levels and results are averaged over 500 randomly generated task instances, but it does not specify explicit training, validation, or test dataset splits for its main evaluation.
Hardware Specification Yes At each deliberation cycle, the time that IGR spends on goal recogition is negligible (less than a millisecond) compared to the time spent on MCTS rollouts (around 4.5 seconds on a Ryzen 9 5900X, with α = 100, β = 10).
Software Dependencies No The paper mentions applying the DQN algorithm but does not provide specific version numbers for software libraries or dependencies used (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes For the goal recogniser, we set δ = 2.5 and η = 0.95. The Qmin parameter of the rollouts (see Algorithm 1) is set to 0.5. For MCTS, we use α = 100, β = 10, c = 2.5.