{"id":294719,"date":"2016-09-19T20:50:07","date_gmt":"2016-09-20T03:50:07","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/?post_type=msr-research-item&p=294719"},"modified":"2018-10-16T22:03:15","modified_gmt":"2018-10-17T05:03:15","slug":"efficient-exploration-dialogue-policy-learning-bbq-networks-replay-buffer-spiking","status":"publish","type":"msr-research-item","link":"https:\/\/www.microsoft.com\/en-us\/research\/publication\/efficient-exploration-dialogue-policy-learning-bbq-networks-replay-buffer-spiking\/","title":{"rendered":"Efficient Exploration for Dialogue Policy Learning with BBQ Networks & Replay Buffer Spiking"},"content":{"rendered":"

When rewards are sparse and action spaces large, Q-learning with \u03f5<\/span><\/span>\u00a0<\/span><\/span> -greedy exploration can be inefficient. This poses problems for otherwise promising applications such as task-oriented dialogue systems, where the primary reward signal, indicating successful completion of a task, requires a complex sequence of appropriate actions. Under these circumstances, a randomly exploring agent might never stumble upon a successful outcome in reasonable time. We present two techniques that significantly improve the efficiency of exploration for deep Q-learning agents in dialogue systems. First, we introduce an exploration technique based on Thompson sampling, drawing Monte Carlo samples from a Bayes-by-backprop neural network, demonstrating marked improvement over common approaches such as \u03f5<\/span><\/span>\u00a0<\/span><\/span> -greedy and Boltzmann exploration. Second, we show that spiking the replay buffer with experiences from a small number of successful episodes, as are easy to harvest for dialogue tasks, can make Q-learning feasible when it might otherwise fail.<\/p>\n","protected":false},"excerpt":{"rendered":"

When rewards are sparse and action spaces large, Q-learning with \u03f5\u00a0 -greedy exploration can be inefficient. This poses problems for otherwise promising applications such as task-oriented dialogue systems, where the primary reward signal, indicating successful completion of a task, requires a complex sequence of appropriate actions. Under these circumstances, a randomly exploring agent might never […]<\/p>\n","protected":false},"featured_media":0,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"footnotes":""},"msr-content-type":[3],"msr-research-highlight":[],"research-area":[13556],"msr-publication-type":[193718],"msr-product-type":[],"msr-focus-area":[],"msr-platform":[],"msr-download-source":[],"msr-locale":[268875],"msr-field-of-study":[],"msr-conference":[],"msr-journal":[],"msr-impact-theme":[],"msr-pillar":[],"class_list":["post-294719","msr-research-item","type-msr-research-item","status-publish","hentry","msr-research-area-artificial-intelligence","msr-locale-en_us"],"msr_publishername":"","msr_edition":"arXiv:1608.05081","msr_affiliation":"","msr_published_date":"2016-08-17","msr_host":"","msr_duration":"","msr_version":"","msr_speaker":"","msr_other_contributors":"","msr_booktitle":"","msr_pages_string":"","msr_chapter":"","msr_isbn":"","msr_journal":"","msr_volume":"","msr_number":"MSR-TR-2016-62","msr_editors":"","msr_series":"","msr_issue":"","msr_organization":"","msr_how_published":"","msr_notes":"arXiv:1608.05081","msr_highlight_text":"","msr_release_tracker_id":"","msr_original_fields_of_study":"","msr_download_urls":"","msr_external_url":"","msr_secondary_video_url":"","msr_longbiography":"","msr_microsoftintellectualproperty":1,"msr_main_download":"","msr_publicationurl":"http:\/\/arxiv.org\/abs\/1608.05081","msr_doi":"","msr_publication_uploader":[{"type":"url","title":"http:\/\/arxiv.org\/abs\/1608.05081","viewUrl":false,"id":false,"label_id":0}],"msr_related_uploader":"","msr_attachments":[{"id":0,"url":"http:\/\/arxiv.org\/abs\/1608.05081"}],"msr-author-ordering":[{"type":"text","value":"Zachary C. Lipton","user_id":0,"rest_url":false},{"type":"user_nicename","value":"jfgao","user_id":32246,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=jfgao"},{"type":"user_nicename","value":"lihongli","user_id":32676,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=lihongli"},{"type":"user_nicename","value":"xiul","user_id":36287,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=xiul"},{"type":"user_nicename","value":"fiahmed","user_id":31810,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=fiahmed"},{"type":"user_nicename","value":"deng","user_id":31602,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=deng"}],"msr_impact_theme":[],"msr_research_lab":[],"msr_event":[],"msr_group":[144931,395930],"msr_project":[377990],"publication":[],"video":[],"download":[],"msr_publication_type":"techreport","_links":{"self":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/294719"}],"collection":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item"}],"about":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-research-item"}],"version-history":[{"count":2,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/294719\/revisions"}],"predecessor-version":[{"id":541645,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/294719\/revisions\/541645"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=294719"}],"wp:term":[{"taxonomy":"msr-content-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-content-type?post=294719"},{"taxonomy":"msr-research-highlight","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-highlight?post=294719"},{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=294719"},{"taxonomy":"msr-publication-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-publication-type?post=294719"},{"taxonomy":"msr-product-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-product-type?post=294719"},{"taxonomy":"msr-focus-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-focus-area?post=294719"},{"taxonomy":"msr-platform","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-platform?post=294719"},{"taxonomy":"msr-download-source","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-download-source?post=294719"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=294719"},{"taxonomy":"msr-field-of-study","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-field-of-study?post=294719"},{"taxonomy":"msr-conference","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-conference?post=294719"},{"taxonomy":"msr-journal","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-journal?post=294719"},{"taxonomy":"msr-impact-theme","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-impact-theme?post=294719"},{"taxonomy":"msr-pillar","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-pillar?post=294719"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}