Change Log for the Multiagent Reinforcement Learning Toolbox ========================================================================= Changes for Version 1.2: Added new world type robotsgw. This is a nicer view over a gridworld. Also supports fully competitive and fully cooperative versions. Supports delayed showing, and delayed slowing down of view updates. Changed episodic_learn to support online viewing of convergence progress. Added reset type support in world_reset. Modified WoLF-Q to support full-state, added switch to disable this Modified teamQ to support social conventions, added switch to disable this ========================================================================= Changes for Version 1.1: Added continuous states and actions support. Hi-level functions made insensitive to the actual type of states and actions. State- and action-space sizes structures changed to accomodate more information such as type of variable, interval bounds for continuous variables etc. 0 action becomes meaningful, noop action is now NaN Added nillearn, nilexplore, and staticact functions, to support static agent behaviors Added support for multiple actions variables per agent Combined agent_act and agent_learn into agent_control Fed actions and rewards into act and explore as well as learn States, (joint) actions and rewards are now column vectors. Consequently, index maps, and state-, action-space sizes are column vectors agent learning parameters fed to agent at construction and not at init_learn Added new robotic arm world. All algorithms in the library currently work only for discrete variables, so none of them work for robotic arm. Currently researching / testing algorithms for continuous spaces.