{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/do-deep-reinforcement-learning-agents-model","title":"Do deep reinforcement learning agents model intentions?","arxiv_id":"1805.06020","date":"2018-05-15","proceeding":null,"authors":["Tambet Matiisen","Aqeel Labash","Daniel Majoral","Jaan Aru","Raul Vicente"],"abstract":"Inferring other agents' mental states such as their knowledge, beliefs and\nintentions is thought to be essential for effective interactions with other\nagents. Recently, multiagent systems trained via deep reinforcement learning\nhave been shown to succeed in solving different tasks, but it remains unclear\nhow each agent modeled or represented other agents in their environment. In\nthis work we test whether deep reinforcement learning agents explicitly\nrepresent other agents' intentions (their specific aims or goals) during a task\nin which the agents had to coordinate the covering of different spots in a 2D\nenvironment. In particular, we tracked over time the performance of a linear\ndecoder trained to predict the final goal of all agents from the hidden state\nof each agent's neural network controller. We observed that the hidden layers\nof agents represented explicit information about other agents' goals, i.e. the\ntarget landmark they ended up covering. We also performed a series of\nexperiments, in which some agents were replaced by others with fixed goals, to\ntest the level of generalization of the trained agents. We noticed that during\nthe training phase the agents developed a differential preference for each\ngoal, which hindered generalization. To alleviate the above problem, we propose\nsimple changes to the MADDPG training algorithm which leads to better\ngeneralization against unseen agents. We believe that training protocols\npromoting more active intention reading mechanisms, e.g. by preventing simple\nsymmetry-breaking solutions, is a promising direction towards achieving a more\nrobust generalization in different cooperative and competitive tasks.","url_abs":"http://arxiv.org/abs/1805.06020v2","url_pdf":"http://arxiv.org/pdf/1805.06020v2.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"do-deep-reinforcement-learning-agents-model","repo_url":"https://github.com/NeuroCSUT/intentions","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"tf","reach":{"status":"unanswered"}}],"tasks":[{"task_slug":"deep-reinforcement-learning","task_name":"Deep Reinforcement Learning"},{"task_slug":"reinforcement-learning","task_name":"Reinforcement Learning"},{"task_slug":"reinforcement-learning-1","task_name":"Reinforcement Learning (RL)"},{"task_slug":"model","task_name":"model"},{"task_slug":"reinforcement-learning-2","task_name":"reinforcement-learning"}],"methods":[{"method_slug":"adam","method_name":"Adam"},{"method_slug":"batch-normalization","method_name":"Batch Normalization"},{"method_slug":"convolution","method_name":"Convolution"},{"method_slug":"dense-connections","method_name":"Dense Connections"},{"method_slug":"experience-replay","method_name":"Experience Replay"},{"method_slug":"maddpg","method_name":"MADDPG"},{"method_slug":"relu","method_name":"ReLU"},{"method_slug":"weight-decay","method_name":"Weight Decay"}],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":null,"mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}