Page 4 of 4 FirstFirst ... 2 3 4
Results 31 to 39 of 39

Thread: pydota2

  1. #31
    Basic Member
    Join Date
    Dec 2016
    Posts
    731
    So I shifted some of my time back to seeing how much information I can extract from protobuf replays for the purposes of RL and the currently github committed Radiant game I recently played (I was the TA) that lasted approx 21 min (slightly less actually) I get the following counts based on animation activities by the 5 players/bots on my team:
    Code:
    [1500] Count:  5921, Perct: 19.55%
    [1502] Count: 20826, Perct: 68.78%
    [1503] Count:  2335, Perct:  7.71%
    [1504] Count:   525, Perct:  1.73%
    [1510] Count:   312, Perct:  1.03%
    [1511] Count:   192, Perct:  0.63%
    [1512] Count:    28, Perct:  0.09%
    [1513] Count:    76, Perct:  0.25%
    [1514] Count:    51, Perct:  0.17%
    [1534] Count:    14, Perct:  0.05%
    This is the mapping of the animation keys to what they mean:
    Code:
            {
                "key": "ACTIVITY_IDLE",
                "value": 1500
            },
            {
                "key": "ACTIVITY_IDLE_RARE",
                "value": 1501
            },
            {
                "key": "ACTIVITY_RUN",
                "value": 1502
            },
            {
                "key": "ACTIVITY_ATTACK",
                "value": 1503
            },
            {
                "key": "ACTIVITY_ATTACK2",
                "value": 1504
            },
            {
                "key": "ACTIVITY_ATTACK_EVENT",
                "value": 1505
            },
            {
                "key": "ACTIVITY_DIE",
                "value": 1506
            },
            {
                "key": "ACTIVITY_FLINCH",
                "value": 1507
            },
            {
                "key": "ACTIVITY_FLAIL",
                "value": 1508
            },
            {
                "key": "ACTIVITY_DISABLED",
                "value": 1509
            },
            {
                "key": "ACTIVITY_CAST_ABILITY_1",
                "value": 1510
            },
            {
                "key": "ACTIVITY_CAST_ABILITY_2",
                "value": 1511
            },
            {
                "key": "ACTIVITY_CAST_ABILITY_3",
                "value": 1512
            },
            {
                "key": "ACTIVITY_CAST_ABILITY_4",
                "value": 1513
            },
            {
                "key": "ACTIVITY_CAST_ABILITY_5",
                "value": 1514
            },
            {
                "key": "ACTIVITY_CAST_ABILITY_6",
                "value": 1515
            },
            {
                "key": "ACTIVITY_OVERRIDE_ABILITY_1",
                "value": 1516
            },
            {
                "key": "ACTIVITY_OVERRIDE_ABILITY_2",
                "value": 1517
            },
            {
                "key": "ACTIVITY_OVERRIDE_ABILITY_3",
                "value": 1518
            },
            {
                "key": "ACTIVITY_OVERRIDE_ABILITY_4",
                "value": 1519
            },
            {
                "key": "ACTIVITY_CHANNEL_ABILITY_1",
                "value": 1520
            },
            {
                "key": "ACTIVITY_CHANNEL_ABILITY_2",
                "value": 1521
            },
            {
                "key": "ACTIVITY_CHANNEL_ABILITY_3",
                "value": 1522
            },
            {
                "key": "ACTIVITY_CHANNEL_ABILITY_4",
                "value": 1523
            },
            {
                "key": "ACTIVITY_CHANNEL_ABILITY_5",
                "value": 1524
            },
            {
                "key": "ACTIVITY_CHANNEL_ABILITY_6",
                "value": 1525
            },
            {
                "key": "ACTIVITY_CHANNEL_END_ABILITY_1",
                "value": 1526
            },
            {
                "key": "ACTIVITY_CHANNEL_END_ABILITY_2",
                "value": 1527
            },
            {
                "key": "ACTIVITY_CHANNEL_END_ABILITY_3",
                "value": 1528
            },
            {
                "key": "ACTIVITY_CHANNEL_END_ABILITY_4",
                "value": 1529
            },
            {
                "key": "ACTIVITY_CHANNEL_END_ABILITY_5",
                "value": 1530
            },
            {
                "key": "ACTIVITY_CHANNEL_END_ABILITY_6",
                "value": 1531
            },
            {
                "key": "ACTIVITY_CONSTANT_LAYER",
                "value": 1532
            },
            {
                "key": "ACTIVITY_CAPTURE",
                "value": 1533
            },
            {
                "key": "ACTIVITY_SPAWN",
                "value": 1534
            },
            {
                "key": "ACTIVITY_KILLTAUNT",
                "value": 1535
            },
            {
                "key": "ACTIVITY_TAUNT",
                "value": 1536
            },
    As you can see, almost 69% of all "work" done by the bots is "moving", 9.4% (1503 & 1504) is on "attacks / right-clicks". As a result, this just re-affirms that pathing/navigation is the largest facet of bot behavior to tackle.

    UPDATE:
    Also, here is the same for the actual ability/items used as reported by "AbilityEvent" protobuf messages:
    Code:
    <ITEMS>
    [1], Count:    36, Perct  2.34%    <-- blink dagger
    [16], Count:     1, Perct  0.07%   <-- branches (yes, I planted a tree with an ironwood branch)
    [34], Count:     5, Perct  0.33%   <-- magic stick
    [36], Count:   112, Perct  7.29%  <-- magic wand
    [38], Count:     3, Perct  0.20%   <-- clarity
    [39], Count:     2, Perct  0.13%   <-- flask
    [40], Count:     6, Perct  0.39%  <-- dust
    [41], Count:    82, Perct  5.34%  <-- bottle
    [42], Count:     8, Perct  0.52%  <-- observer ward
    [44], Count:    25, Perct  1.63%  <-- tango
    [45], Count:     1, Perct  0.07%   <-- courier (1 courier)
    [46], Count:   105, Perct  6.84%  <-- tpscroll
    [48], Count:    10, Perct  0.65%   <-- BoTs
    [63], Count:     2, Perct  0.13%
    [79], Count:     4, Perct  0.26%
    [92], Count:    16, Perct  1.04%  <-- urn of shadows
    [100], Count:     2, Perct  0.13%
    [102], Count:     4, Perct  0.26%
    [131], Count:     3, Perct  0.20%
    [172], Count:     8, Perct  0.52%
    [180], Count:    42, Perct  2.73%  <-- arcane boots
    [185], Count:     6, Perct  0.39%
    [188], Count:     4, Perct  0.26%
    [216], Count:     1, Perct  0.07%
    [237], Count:     1, Perct  0.07%
    [241], Count:     1, Perct  0.07%
    [254], Count:    11, Perct  0.72%   <-- glimmer cape
    [257], Count:     2, Perct  0.13%
    
    <ABILITIES>
    [5040] Count:    66, Perct:  4.30% -- lina_dragon_slave
    [5041] Count:    54, Perct:  3.52% -- lina_light_strike_array
    [5043] Count:     9, Perct:  0.59% -- lina_laguna_blade
    [5094] Count:    34, Perct:  2.21% -- sven_storm_bolt
    [5096] Count:    22, Perct:  1.43% -- sven_warcry
    [5097] Count:    11, Perct:  0.72% -- sven_gods_strength
    [5182] Count:    47, Perct:  3.06% -- faceless_void_time_walk
    [5185] Count:     7, Perct:  0.46% -- faceless_void_chronosphere
    [5194] Count:    65, Perct:  4.23% -- templar_assassin_refraction
    [5195] Count:    44, Perct:  2.86% -- templar_assassin_meld
    [5197] Count:    51, Perct:  3.32% -- templar_assassin_psionic_trap
    [5198] Count:    42, Perct:  2.73% -- templar_assassin_trap
    [5226] Count:    39, Perct:  2.54% -- dragon_knight_breathe_fire
    [5227] Count:    19, Perct:  1.24% -- dragon_knight_dragon_tail
    [5229] Count:    10, Perct:  0.65% -- dragon_knight_elder_dragon_form
    [5263] Count:    49, Perct:  3.19% -- omniknight_purification
    [5264] Count:    17, Perct:  1.11% -- omniknight_repel
    [5266] Count:     6, Perct:  0.39% -- omniknight_guardian_angel
    [5353] Count:    32, Perct:  2.08% -- spirit_breaker_charge_of_darkness
    [5354] Count:    23, Perct:  1.50% -- spirit_breaker_empowering_haste
    [5356] Count:    12, Perct:  0.78% -- spirit_breaker_nether_strike
    [5548] Count:    72, Perct:  4.69% -- bristleback_viscous_nasal_goo
    [5549] Count:   104, Perct:  6.77% -- bristleback_quill_spray
    [5637] Count:    49, Perct:  3.19% -- oracle_fortunes_end
    [5638] Count:     5, Perct:  0.33% -- oracle_fates_edict
    [5639] Count:    41, Perct:  2.67% -- oracle_purifying_flames
    [5640] Count:     8, Perct:  0.52% -- oracle_false_promise
    [5651] Count:    21, Perct:  1.37% -- winter_wyvern_arctic_burn
    [5652] Count:    36, Perct:  2.34% -- winter_wyvern_splinter_blast
    [5653] Count:    21, Perct:  1.37% -- winter_wyvern_cold_embrace
    [5654] Count:     6, Perct:  0.39% -- winter_wyvern_winters_curse
    [5691] Count:    11, Perct:  0.72% -- faceless_void_time_dilation
    Last edited by nostrademous; 12-06-2017 at 06:59 AM.

  2. #32
    Basic Member
    Join Date
    Mar 2012
    Posts
    2,013
    You also get the target of item or ability usages in the protobuff, correct? The reason I am asking is, I was wondering how it would serialize a hero eating a normal tree vs an iron tree vs a sprout tree.
    I ask because atm I don't think there is a way to target "non-map" trees.

    Also, since I assume it's an integer target, keep in mind that you'll have to always track the item to not assume the int is a unit vs a tree.
    I am mostly thinking of GetBotByHandle() calls here.
    Last edited by The Nomad; 12-07-2017 at 11:58 AM.
    Explanations on the normal, high and very high brackets in replays: here, here & here
    Why maphacks won't work in D2: here

  3. #33
    Basic Member
    Join Date
    Dec 2016
    Posts
    731
    No, the target is not provided as far as I know. Only for EventDamage messages.

  4. #34

    Interesting

    +1
    Going through the approach. A bit of novice myself on ML, will try to help

  5. #35
    Basic Member
    Join Date
    Dec 2016
    Posts
    731
    Back from the Holidays. Starting to submodule and improve a Dota Simulator (which will be necessary) started by lenLRX into the project.

  6. #36
    @nostrademous Hey, awesome work! I'd love to contribute to this. I've been trying out several policy gradient and evolutionary techniques for toy problems, as well as for robot soccer (https://github.com/korroktheslavemaster/rl-algos super hacky code). From what I gather, your client/server model will not work in custom games since CMsgBotWorldState is not broadcasted in that mode (atleast not when I tried, maybe you can help?)? And without custom games how are you planning to make different environments (creep block only, last hit only etc.)?

  7. #37
    Basic Member
    Join Date
    Dec 2016
    Posts
    731
    @theslavemaster I plan on supporting custom games eventually (and will default to using the polling CreateRemoteHTTPPost() to pass data), but currently working on implementing Dota2 simulator support (with help from @lenLRX) to allow for parallelized non-GUI learning of things that can easily be simulated.

  8. #38
    OK, thanks. I wanted to do episodic training, where I can reset the environment (eg. despawn creeps, reposition heroes) to start a new episode, but I don't think it would be possible with this right now I hope that functionality is present in the simulator?

  9. #39
    I went through the simulator code briefly. I feel like Dota's physics engine will be required for a lot of interesting tasks (creep blocking, even last hitting for melee heroes) and will definitely be required for more complex environments. I'm not sure if it's even feasible to get that into the simulator... what specific environments do you have in mind now which you're adding to the simulator?

Tags for this Thread

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •