Dr. Matthew Crosby
I am currently working on the Kinds of Intelligence project part of the Leverhulme Centre for the Future of Intelligence (CFI) at Imperial College London.
I am particularly interested in computational philosophy (of mind); implementing and testing AI systems to explore philosophical questions about the nature of intelligence. I am also interested in consciousness, though believe in an incremental approach based on improving our understanding of the space of possible minds.
I specifically want to be able to improve our answer to the following:
For any given system, what kind of mind does that system have (if any)?
My current research is focused on linking active inference with deep learning in embedded systems. I am particularly interested in learning algorithms that incorporate how actions change the inputs to a system. I am also interested in carving out a unique causal role for consciousness and what (if anything) conscious thought offers a system over purely unconscious processing.
A (very rough) overview of my current research plan can be found here.
These publications are mainly from a previous life where I worked in multiagent planning and robotics.
Please also see my Google Scholar page
- SkiROS: A skill-based robot control architecture on top of ROS Francesco Rovida, Matthew Crosby, Dirk Holz, Athanasios S. Polydoros, Bjarne Grossmann, Ronald P. A. Petrick and Volker Kruger, Robot Operating System (ROS) Springer, 2017 [Book Chapter, link]
- Integrating Mission and Task Planning in an Industrial Robotics Framework Matthew Crosby, Francesco Rovida, Volker Krueger and Ron Petrick, International Converence on Automated Planning and Scheduling (ICAPS) 2017 [pdf]
- A Vertical and Cyber-Physical Integration of Cognitive Robots in Manufacturing Volker Krueger et al, Proceedings of the IEEE Volume:104, Issue: 5 2016 [link] Describes the robot architecture created for the STAMINA project.
- ADP: an Agent Decomposition Planner CoDMAP 2015 Matthew Crosby, ICAPS Proceedings of the Competition of Distributed and Multi-Agent Planners (CodMAP) 2015 [pdf]>
- A Single-Agent Approach to Multiagent Planning Matthew Crosby, Anders Jonsson, and Michael Rovatsos, 21st European Conference on Artificial Intelligence (ECAI) 2014 [pdf]
- Improving Planner Performance in Grid Worlds with Macro Actions Matthew Crosby and Ronald P. A. Petrick, The 9th International Workshop on Cognitive Robotics (CogRob) at the 21st European Conference on Artificial Intelligence (ECAI) 2014 (Short Paper) [pdf]
- Temporal Multiagent Planning with Concurrent Action Constraints Matthew Crosby, R.Petrick, Distributed and Multiagent Planning workshop (DMAP) at the International Conference on Automated Planning and Scheduling (ICAPS) 2014 [pdf]
- Multiagent Classical Planning Matthew Crosby, PhD Thesis 2014 [pdf] Mainly about solving classical planning problems quickly by decomposing into subproblems (that resemble separate agents). Also contains some work on centralised planning for self-interested agents and concurrent action planning.
- Automated Agent Decomposition for Classical Planning Matthew Crosby, M.Rovatsos and R.Petrick, International Conference on Automated Planning and Scheduling (ICAPS) 2013 [pdf] Introduction to the ADP algorithm which automatically decomposes planning problems into separate agents and then employs a agentified variation of the FF heuristic to quickly find plans.
- Heuristic Multiagent Planning with Self-Interested Agents Matthew Crosby and M.Rovatsos, International Conference on Autonomous Agents and Multiagent Systems (AAMAS) 2011 (Extended Abstract) [pdf] Contains a centralised planning algorithm that outputs stable plans for self-interested agents on a restricted class of planning problems (Safe-CoPGS). A more detailed look can be found in my PhD thesis (though this topic did not become the main focus of the work).
- Evolving a Roving Eye for Go Matthew Crosby, Master's Thesis 2008 [pdf] In which the results of a paper aiming to create a go playing roving eye capable of 'seeing' only a 3x3 grid of the go board at a time were attempted to be reproduced. This could have been subtitled: "On why using a small input space for a global problem and testing your program against a deterministic games player is a bad idea". Or: "Who knew a 19 x 19 x 48 image stack would be a feasible sized input to a neural network 5 years down the line? (Alphago)"
ADP Planner (planner with best coverage in CodMAP 2015 competition) - including the snapshot of Fast-Downward it was built for - can be found here. After compiling, run with the option --heuristic 'hff=adp(cost_type=1)' --search 'lazy_greedy(hff, preferred=hff)'. See Fast Downward site for required libraries in case of installation issues.
Python ma-PDDL Parser ppp.py: for translating multiagent pddl with concurrent action constraints to temporal planning problems (see DMAP 2014 paper [pdf]). An updated version was used in the CodMAP 2015 competition (which may be more useful to you), see here for information and contacts.
Notes for the computational neuroscience reading group held in Edinburgh (Jan-April 2017) can be found here.