Anyone in the fediverse attending #icaps in #prague?

#icaps2023 #planning #automatedplanning #AIPlanning

@billjanssen Agreed.

Unfortunately for me (and us at #PARC) we love to be in the scientific cracks and do inter-disciplinary research. So we are constantly fighting this war of oh this is not #HCI, not #AI enough, this sorta looks like #AIPlanning but not, this is more transportation than #AI.

Thankfully, the journals are more creative.

On the other hand, some #LLM papers claim #AIPlanning while not really solving the "learning how to plan" problem. These papers were published at exclusively #NLP venues and had no (apparent) input from #AIPlanning or #Agents communities.

Finally, #AI publishing is operating with this rule:

if #ML #DL -> can do magical things.
If not #ML #DL -> why aren't you doing #ML #DL or what about this other #ML method claiming to solve this other problem.

#AI #ML research/publishing operates in silos - to the detriment of making progress.

Our #IJCAI submission on #OpenWorldLearning #OWL was rejected for good and bad reasons.

The bad reason is: "this is not just planning but also something similar to reinforcement learning".

Guess what - that is the point of our research! We are trying to close the gap between designed #AIPlanning systems and adaptive #Learning systems. It is a super-hard gap to push #AI #ML algorithmic research in.

Common wisdom in #AI and #ML is that #AIPlanning methods cannot deal with continuous state and action spaces.

Subverting these expectations - presenting our recent paper on how a planning agent can play #AngryBirds!

And, no #DL #DQN systems cannot play these games yet AND take so much data to learn to play a single level.

#Planning #Reasoning #KRR FTW!!

https://arxiv.org/abs/2303.16967

January has been an exciting month for #AI #ML fundamental research at #PARC.

Our work on making #AIPlanning methods work/learn in an #OpenWorld -will be presented at #AAMAS2023 as well as at #ICAPS2023. AND, an #AIJ article is under works.

#OpenWorldLearning is a new challenge - the environments introduce novelties while the agent is operating in the world. The agent must detect, characterize, and accommodate novelties during run time. This research is a part of #DARPA #SAILON program

1998. SharedPlan theory of discourse culminates in research with actual software implementations of written-dialogue-capable agents in GUI software: https://doi.org/10.1007/978-94-017-1118-0_4

It’s like the old #Clippy from Microsoft Office, but actually focussed on your shared goals, and figures out what to say instead of being too pre-programmed.

I feel that goal-oriented approaches, using #AIPlanning, are still relevant to produce dialogues along with physical actions, in a complement to #LLM-style dialogues.

At the meeting of the French #robotics research group #GdRRobotique, I liked Nick Hawes’ STRAND project about long-term #HRI. For 3 months they had an interactive robot:
- explore to learn times and locations more eager to lead to interactions into an #MDP
- assess the risk of requiring an interaction to move on, e.g. when it needed someone to open a door
- exploit by maximising the number of interactions per day, using a tool called PRISM #AIPlanning
https://arxiv.org/pdf/1604.04384.pdf

FailRecOnt - Failure Recovery #Ontology : http://hdl.handle.net/2117/357471
A framework for reasoning about action failures in #robotics, based on the IEEE Ontology for Autonomous Robots (via the DUL ontology), that deduces causes of failures and suggest actions for recovery.

Besides recovery, it demonstrates how #AIPlanning from symbolic knowledge is #explainableai.